id
stringlengths 24
24
| idx
int64 0
402
| paragraph
stringlengths 106
17.2k
|
|---|---|---|
65ef10259138d23161364879 | 9 | In order to account for the effect of water quality, surface water and groundwater were modelled based on parameters and assumptions derived from literature. E. coli was selected as the indicator microbe to evaluate the level of contamination in the study. The algorithms for the operation and maintenance requirements for each technology were designed to achieve a consistent log reduction and disinfection efficacy of 3 logs (at a minimum) for all systems evaluated. To achieve the set level of efficacy for each system, a comprehensive assessment was conducted to determine the necessary capital materials and consumables based on raw water quality. This assessment had implications for both cost and environmental considerations. Therefore, the different water quality parameters for groundwater and surface water sources serve as the contextual parameters modelled into the systems. For instance, groundwater source will most likely have higher hardness due to water dissolves minerals as it moves through rocks. Both waters have 1,500 -2,500 CFU⋅100 mL -1 of E. coli. A characteristic groundwater had a turbidity of 1 -10 NTU and hardness of 60 -120 mg⋅L -1 as CaCO3, and characteristic surface water had a turbidity of 10 -30 NTU and hardness of 0 -60 mg⋅L -1 as CaCO3. |
65ef10259138d23161364879 | 10 | We accounted for capital, operation and maintenance, and energy costs. All costs were normalized to the economic indicator of USD•cap -1 •yr -1 . Specifically, capital cost covered all the purchases that the units required at start (e.g., housing for the UV systems, water storage bottles) while operation and maintenance accounted for cost estimates of all consumables materials and parts that require periodic replacements (e.g., NaClO, lamps, AgNP coating). |
65ef10259138d23161364879 | 11 | Energy cost requirements were accounted for depending on the electricity need of each unit. It is notable that this requirement does not apply to units without electricity use (i.e., POU chlorination and AgNP CWFs). The initial step of TEA involves identifying the specific objective for cost assessment, determining the components comprising the technology, and identifying the various factors that contribute to overall cost (e.g., cost of UV lamp, labor cost). The next step entails data compilation of on the cost associated with each material and determining the frequency at which such costs will be applicable in cases involving replaceable parts. It is important to note that capital costs are also spread out through the analysis period (5 years baseline period). Discounted cash flow analysis was applied to account for future value of money over the technology's lifespan with a 5% discount rate on average. Subsequently, the following step involves identifying and considering capital costs associated with construction, operation, and maintenance over the entire duration of the analysis. The cost analysis was designed to account for impacts of water quality from each unit while achieving necessary disinfection efficacy. |
65ef10259138d23161364879 | 12 | LCA of the POU technologies encompassed impacts from capital inputs, operational activities, maintenance requirements, and energy consumption. Life cycle greenhouse gas emission impact data was obtained from EcoInvent v3.9 database considering all materials and consumables in each unit, and global warming potential (GWP) was selected as the environmental sustainability indicator through the U.S. EPA's TRACI (Tool for the Reduction and Assessment of Chemicals and Other Environmental Impacts) method. The LCA methodology employed in this study followed several key steps. Firstly, the goal and scope were established to track the environmental impacts associated with both the capital inputs and the operation and maintenance requirements of the analyzed POU technologies. The inventory analysis was used to account for all the materials and their respective weights (in kg) and other relevant parameters (such as the number of UV lamps utilized) in each POU system. The impact assessment phase incorporated the GWP for the identified parameters and materials. |
65ef10259138d23161364879 | 13 | Uncertainty was incorporated into all assumptions and data for each parameter by introducing a range of 5-25% of uncertainty distribution depending on the data availability and level of confidence. The incorporated uncertainties capture variation in the values for all the data points, e.g., fluctuation in materials cost and impacts. To address and quantify uncertainty, a total of 10,000 Monte Carlo simulations were conducted. Sensitivity analysis was performed to determine factors and parameters that are key drivers to changes in system's cost and environmental impacts. Specifically, we used the Spearman's rank correlation coefficients to measure and analyze the sensitivity of individual parameters for all units. Here, we report the absolute value for the top five Spearman's rank correlation coefficients (> |0.05| and p-value < 0.05) for total cost and GWP for each technology in each water. |
65ef10259138d23161364879 | 14 | The baseline assumption in this study was that each POU technology would be utilized for a duration of 5 years. However, to gain deeper insights, the analysis further examined the impact of adopting POU technologies for different lifetimes. Depending on the context, these technologies may be deployed for a relatively short period (e.g., after extreme weather events cause interruption of a centralize water supply) or a longer period (e.g., as a primary treatment method in underserved communities). The performance of each technology was simulated by setting the usage period to 1, 2, 5, 10, and 15 years. The design and process algorithms for each technology were adjusted accordingly to account for the change in the usage period to obtain the net cost and net GWP associated with different lengths of technology adoption. |
65ef10259138d23161364879 | 15 | To provide insight on deploying the four POU technologies across the world, a contextual analysis was performed to assess the implications of contextual parameters specific to the deployment site. Demographic (household size), water quality (E. coli, turbidity, hardness), and energy (electricity cost and GWP characterization factor) data were collected from ten different communities. These communities include two from Africa (Kampala, Uganda; Limpopo, South Africa 38 ), two from Asia (Gunungkidul, Indonesia; 39 Panobolon Island, Philippines 40 ), four from North America (Colonias, United States; Navajo Nation, United States; Les Anglais, Haiti ; Oaxaca, Mexico ), and two from South America (Santa Cruz, Bolivia 45 ; Antioquia, Colombia 46 ). The collected data were then used in TEA and LCA to obtain location-specific cost and GWP. |
65ef10259138d23161364879 | 16 | For the groundwater, the POU chlorination system was found to have the lowest cost with a net cost of 0.09 [0.05 -0.020] USD•cap -1 •yr -1 (median [5 th -95 th hereinafter]; Figure ). The next lowest cost was AgNP CWF at 0.43 [0.31 -0.65] USD•cap -1 •yr -1 . UV (mercury) lamp had a net cost of 4.96 [3.04 -10.18] USD•cap -1 •yr -1 , and the highest net cost was for UV LED which was 18.32 [10.08 -42.49] USD•cap -1 •yr -1 . The cost-effectiveness of POU chlorine treatment can be attributed to the utilization of simple and affordable materials like 20 L jerrycans and WaterGuard (NaClO) bottles, which are available at a cost ranging from 0.08 to 0.33 USD per bottle. The low cost for AgNP CWF can be attributed to the low cost of capital materials and production along with low-cost requirements for operation and maintenance. The only consumable for AgNP CWFs is the AgNP recoating which is not as frequent compared to the POU chlorination system that relies strictly on more affordable consumable NaClO. In contrast, UV systems employing mercury lamps and UV LEDs involve relatively higher costs due to requirement for more expensive materials and the consumption of electricity during operation. When comparing the two UV systems, it is observed that UV LEDs are generally more expensive than mercury lamps. However, UV LEDs offer a longer lifespan and have lower electricity requirements compared to traditional mercury lamps. |
65ef10259138d23161364879 | 17 | In the case of surface water, the net cost followed the same order groundwater, but the specific cost estimates were higher for each technology. The net cost for surface water, from lowest to highest, were as follows: POU chlorination (0.11 [0.07 -3.65] USD•cap -1 •yr -1 ), AgNP CWF (0.52 [0.36 -0.90] USD•cap -1 •yr -1 ), POU UV mercury lamp (5.96 [3.57 -13.40] USD•cap - 1 •yr -1 ), and UV LED (23.97 [12.19 -49.97] USD•cap -1 •yr -1 ). The higher operation and maintenance cost associated with surface water is primarily due to the need of replaceable parts or consumables. This cost increase is more significant for chlorination and only slightly higher for AgNP CWFs. Specifically, due to the higher turbidity level, the surface water required a higher dose (doubling the dose) of NaClO for POU chlorination. Similarly, the increased in turbidity required more frequent recoating for the AgNP CWFs. It is worth noting that the increase in cost for the AgNP CWF is marginal. The turbidity in surface water also leads to an increased electricity run time for UV mercury lamps and UV LEDs, resulting in higher electricity costs and more frequent lamp replacements. However, these additional costs have minimal impact on the overall costs of the UV mercury lamps and UV LEDs. Overall, the higher operation and maintenance costs associated with surface water result in higher net costs for deploying POU technologies when treating raw water with similar water characteristics to groundwater. |
65ef10259138d23161364879 | 18 | Regarding environmental impacts, for groundwater, AgNP CWF technology exhibited the lowest overall GWP, estimated to be 0.04 [0.03 -0.07] kg CO2 eq•cap -1 •yr -1 , which was followed by POU chlorination with had an estimated GWP of 0.12 [0.07 -0.28] kg CO2 eq•cap -1 •yr -1 (Figure ). The UV LED had a higher GWP of 1.51 [0.84 -3.49] kg CO2 eq•cap -1 •yr -1 , while the UV mercury lamp technology had the highest GWP, estimated at 2.55 [1.50 -5.29] kg CO2 eq•cap -1 •yr -1 . For both UV systems, the impact on GWP from capital materials was greater than that from operation and maintenance, which mainly consisted of electricity consumption and lamp replacement. However, the POU chlorination system was also influenced by operation and maintenance costs due to the need for consumable NaClO. |
65ef10259138d23161364879 | 19 | Overall, the cost and environmental impacts of these POU disinfection technologies can be directly influenced by water quality. Turbidity in treated water necessitates increased consumables for effective disinfection across all technologies. These consumables can have a direct influence on overall sustainability. Understanding the capital, operation, and maintenance requirements can help inform the deployment of these POU technologies in various contexts. For instance, chlorination relies heavily on the NaClO supply chain, while the UV systems require a readily available electricity source. This level of analysis reveals that characteristics of the source water can significantly impact sustainability and the specific requirements of each technology offer different opportunities for deployment. |
65ef10259138d23161364879 | 20 | Overall, the key drivers were similar for the disinfection of groundwater and surface water (Figure and Figure ). The discount rate was found to have noticeable influence the cost of all four of the technologies. For POU chlorination, the assumptions that influenced cost were the dose of NaClO and the chlorination container cost (Figure ). This outcome is expected since NaClO is the primary consumable in this technology, and the container is the only capital requirement. In the case of AgNP CWF, the key drivers were labor cost, bucket cost, and spout cost. For water type 2, with AgNP CWF, the key drivers were AgNP loading rate, labor cost, discount rate, bucket cost, and lid cost. |
65ef10259138d23161364879 | 21 | The key drivers of GWP for each technology are presented concerning the two water types (Figure and Figure ). For POU chlorination, the key drivers of GWP were the weight of the polyethylene (PE) container and PE characterization factor. The 20 L plastic container used in the system had a significant impact on LCA of the POU chlorination system for both water types as it is a capital component of the system. |
65ef10259138d23161364879 | 22 | Regarding AgNP CWF, the key drivers were the PE characterization factor, the PE in container, and the AgNP loading. The AgNP loading refers to the concentration of AgNP on the CWF based on the mass of AgNP applied per filter. The component with the highest influence on environmental impact of AgNP CWF system was the plastic bucket that holds the water that filters through the CWF. The AgNP coating had more influence on surface water due to the shorter AgNP lifespan which results in more frequent AgNP recoating in response to the higher turbidity of the water. |
65ef10259138d23161364879 | 23 | For UV mercury lamp system, the key drivers were UV mercury lamp lifespan, UV mercury lamp impact factor, aluminum impact factor, aluminum foil weight, and PE from storage. For water type 2, the key drivers were UV mercury lamp characterization factor, PE characterization factor, UV mercury lamp lifespan, and aluminum weight. The key drivers of the UV mercury lamp system primarily revolved around capital requirements and lamp replacement. The UV lamps are key drivers of GWP and can be attributed to the lamp's mercury content and the release of mercury into the environment during disposal. For the UV LED system, the key drivers were LED characterization factor, PE characterization factor, stainless steel characterization factor, LED lifespan, stainless steel weight, UV LED weight, and PE weight. Both UV systems were impacted by the lifespan of the lamps and LEDs, as lamp replacement is necessary over time. |
65ef10259138d23161364879 | 24 | Overall, the results from the sensitivity analysis highlight the influence of assumptions on the financial and environmental sustainability of the POU technologies. The identification of key drivers can also guide technology developers in areas to focus on for research and improvement. For instance, when deploying POU chlorination using WaterGuard or similar products as a source of NaClO, then the desired dose of NaClO will be an important factor to consider while adjusting for cost and environmental impacts. The cost of the AgNP CWF is primarily impacted by labor to manufacture the filters, suggesting that exploring mass production methods may further reduce costs. Lowering the unit cost is a key area for improving the cost of both UV systems. The negative Spearman's rank correlation of GWP impact of lamp and LED lifespan on GWP indicates that enhancing lifespan can increase environmental sustainability. These key drivers can provide a potential pathway for technology developers and manufacturers to improve the sustainability of these POU technologies. |
65ef10259138d23161364879 | 25 | The adoption lifetime refers to the expected duration in years for which a household is likely to use a specific POU technology. In some cases, a POU technology may be deployed for shortterm interventions, such as disasters relief efforts, or for long-term usage and treatment interventions, particularly in developing regions. For each POU technology, the study analyzed the cost and environment impacts associated with adoption and usage periods ranging from 1 to 15 years. Across all POU technologies, a consistent trend was observed: as the adoption lifetime increased, the yearly per capita cost and environmental impact decreased. This overall trend indicates that POU technologies exhibit greater sustainability with long-term adoption and usage. Long-term adoption is advantageous because it allows for the spreading out of costs and environmental impact over a greater number of years, as opposed to investing in a technology and only using it for a short period. However, it is important to note that the extent of cost and environmental impact reduction with longer lifetimes varies significantly among the different POU technologies. |
65ef10259138d23161364879 | 26 | The net costs and GWP for all four technology during a 1-year adoption period (Figures ) were used to normalize results to their respective median from different adoption periods (Figures ). For POU chlorination, all values were lower with longer-term adoption. In the case of short-term adoption (1 year), the POU chlorination system had a median net cost of 0.42 USD•cap -1 •yr -1 and median net GWP of 0.62 kg CO2 eq•cap -1 •yr -1 . However, for long-term adoption (15 years) the net cost decreased to 9.31% of the 1-year adoption cost. On the other hand, the GWP for 15 years decreased to 6.67% of the 1-year adoption scenario. These reductions in cost and GWP with longer adoption periods are due to the distribution of capital requirements associated with the 20 L jerry can over the extended lifetime of the system. It does appear that both indicators level out at higher adoption periods, which can be attributed to the continuous need for consumables (i.e., NaClO) to run the system. The AgNP CWF system exhibited the lowest GWP and the second lowest costs compared to all other technologies across the entire range of adoption lengths (Figure ). Specifically, for a 1-year adoption term, the estimated net cost was 3.31 USD•cap -1 •yr -1 and the net GWP was 0.33 kg CO2 eq•cap -1 •yr -1 . Both the net cost and GWP significantly decreased as the adoption period increased from 1 to 5 years. At a 15 years adoption term, the estimated net cost and GWP were 9.31% and 7.92% of the 1-year adoption, respectively (Figures ). Therefore, for both short-term and long-term adoption, the AgNP CWF system appears to be a viable option. This system has the potential to be the most sustainable choice, considering both cost and environmental impacts. |
65ef10259138d23161364879 | 27 | Similarly, both UV systems had pronounced decline in cost and a moderate decline in GWP with an increase in adoption lifetime. This finding can be attributed to the higher capital cost requirements associated with these advanced systems. In the case of the UV mercury lamp system, a 1-year adoption period was associated with a net cost of 24.59 USD•cap -1 •yr -1 and a GWP of 7.28 kg CO2 eq•cap -1 •yr -1 (Figure ). However, the cost decreased significantly after approximately 5 years, with 15 years adoption period resulting in a net cost of 9.74% of the 1-year adoption cost (Figure ). On the other hand, the 15-year adoption period resulted in 61.81% of the 1-year adoption GWP (Figure ). This only moderate reduction can be attributed to the GWP from the mercury lamps that require replacement throughout the adoption period. The UV LED system had the highest cost of all the POU technologies over the entire range of adoption periods (Figure ). A 1-year adoption was associated with net cost of 64.92 USD•cap -1 •yr -1 , while a 15-year adoption yielded a net cost was 8.95% of the 1-year adoption cost (Figure ). The GWP for a 1-year adoption was 4.15 kg CO2 eq•cap -1 •yr -1 and 32.10% of the 1-year adoption GWP (Figure ). The moderate reduction in GWP with adoption period for the UV LED system can also be attributed to the required lamp replacement. Overall, the drastic reduction in costs versus moderate reduction in GWP with adoption period for the UVbased system present tradeoffs in their adoption. These results suggest that long-term adoption is the preferred approach when considering the costs of UV systems. percentiles are plotted with the dashed lines. Note that the household size was set to 4 people to focus on how adoption period can influence cost and environmental impacts. |
65ef10259138d23161364879 | 28 | As communities across the world are characterized by their unique economic, environmental, and social situations, location-specific parameters beyond technology specifications may also have substantial impacts on overall sustainability. To explore these potential implications, ten communities from four continents were included in a contextual analysis where TEA and LCA of the four POU technologies were performed with community and/or region-level demographic, water quality, and energy data (Table ). Consistent previous results, POU chlorination and AgNP CWF had much lower costs and GWP than UV lamp and UV LED, regardless of the deployment site (Figure ). However, different trends were observed depending on the specific type of technology. For POU chlorination and AgNP CWF where the capital cost and construction of the equipment were cost and environmental impact drivers, per capita cost and GWP were found to negatively correlated to the size of the household, with Colonias in the United States (household size of 6.48) and Santa Cruz, Bolivia (household size of 5) on the lower end and Gunungkidul, Indonesia, Limpopo, South Africa, Navajo Nation, United States, and Oaxaca, Mexico on the higher end (household size < 4). For the two UV technologies, however, different trends were found for cost vs. GWP, which were also correlated to both the water quality and the electricity profile of the community. For example, for Gunungkidul, Indonesia which had the highest costs and GWP for POU chlorination and AgNP CWF (smallest household), thought it still had the highest cost for UV lamp and UV LED, the GWP of these two UV systems were lower than those of Limpopo, South Africa (highest among all), Kampala, Uganda, and Les Anglais, Haiti. This change in trend was due to Gunungkidul's comparably lower turbidity (0.36 NTU 39 , would allow longer equipment lifetime and less electricity consumption) and a cleaner (0.687 kg CO2eq⋅kWh -1 , 48 vs. 1.014 kg CO2eq⋅kWh -1 for Limpopo, South Africa 49 ) grid in Indonesia. Notably, electricity was not identified as a driver for GWP in the sensitivity analysis, likely due to the narrower ranges considered previously (0.52 kg CO2 eq⋅kWh -1 to 0.87). Finally, it should be noted that as these communities have very limited income, cost is nonetheless still likely to be the largest hurdle for the adoption of the UV technologies, which were found to be orders of magnitudes higher than POU chlorination and AgNP CWF. |
65ef10259138d23161364879 | 29 | The QSD framework was leveraged in this study to compare the performance of POU technologies in terms of cost and environmental impacts. Based on the economic analysis, the POU chlorination system had the lowest net cost, while the UV LED system had the highest net cost, considering the baseline general assumptions. In terms of environmental impacts, the AgNP CWF system exhibited the lowest GWP, whereas the UV mercury lamp system had the highest environmental impacts, again based on the baseline general assumptions. If the motivation for selecting a technology is affordability, especially in low-income areas, POU chlorination would be appropriate for short-term adoption, while AgNP CWF may be more suitable for long-term adoption. On the other hand, if GWP is the deciding factor for selecting a technology, AgNP CWF would be appropriate based on the reported low environmental impacts, as revealed in this study. It is worth noting that the AgNP CWF is user-friendly; however, the process of recoating the AgNPs onto the CWF will require an expert assistance, compared to POU chlorination, which households can easily use without needing expert involvement. On the UV systems, our findings indicate that UV LED had the higher cost under all adoption periods, but its GWP was lower compared to UV mercury lamps due to the disposal phase of the mercury of the lamps. However, due to the electricity demand, both UV systems would be less effective in regions where electricity supply is not adequate or unavailable. |
65ef10259138d23161364879 | 30 | With regard to water quality, owing to the increased requirement for replaceable components or consumables to achieve effective disinfection, more turbid water would lead to higher net GWP for all POU technologies, underscoring the importance of technology developers to evaluate the impact of different water sources on the sustainability of their systems. Further, the change in water quality would also propagate effects on sustainability drivers, such as the case where NaClO dosage needs to be adjusted to align with the turbidity of the raw water. Moreover, this study also revealed the significance of considering locationspecific parameters for technology deployment. Using data specific to ten communities across the world, we showed that the significant variations in the cost and GWP of these four POU technologies. |
65ef10259138d23161364879 | 31 | Meanwhile, it is important to acknowledge that the results and findings in this study are under a set of assumptions derived from manufacturer recommendations, published reports, and scientific papers. The specific results and outcomes can vary depending on the changes in key assumptions and parameters that drive sustainability. Moreover, the inclusion of additional decision variables, contextual parameters, and technological parameters may yield different outcomes. For instance, factors such as the cost of water transportation from source or the energy required for groundwater pumping may have an impact. Incorporating these additional parameters or modifying existing ones in future analyses can yield more contextspecific and informed results. Thus, this study can serve as a foundation for future researchers and entities interested in understanding the relative sustainability of different POU technologies. Finally, while this this study focused on four selected POU technologies, the framework employed can be extended to explore other POU technologies including novel and emerging ones, that need to be evaluated prior to deploying. Therefore, this study has the potential to help inform research, development, and deployment of POU disinfection technologies considering decision variables, technological parameters, and contextual parameters. |
66b51a695101a2ffa89c19fe | 0 | The consideration of spin-orbit coupling is crucial for photochemistry , enabling, even for light atoms, the coupling of energetically close lying excited states of different spin multiplicity and thus resulting in a splitting of spectral lines and inducing spin-forbidden transitions. The latter play an emerging role in materials science, e.g. for light-induced excited-state spin trapping or upconversion processes. Sophisticated quantum-mechanical descriptions of spin-orbit coupling (SOC), the interaction between orbital angular momenta and spin angular momenta of electrons, based on the Dirac-Coulomb-Hamiltonian comprise four-or two-component approaches combined with various electronic-structure methods. Computational costs of excisting relativistic schemes are however at least one order of magnitude larger than comparable non-relativistic or scalar-relativistic computations, motivating the use of perturbative SOC ansätze . Relying on (quasi)degenerate perturbation theory , spin-free states are thereby provided within a non-relativistic framework and coupled via approximate spin-orbit Hamiltonians as the perturbation. A perturbative treatment of SOC was found to be reliable as long as SOC interactions are of small magnitude, which was found to be the case for molecules not containing 6p-or 7p-block elements. As highlighted by Cheng et al, it were perturbative SOC ansätze that pioneered the original description of intersystem crossing and current implementations for the dynamic description of ISC relying on surface hopping nonadiabatic molecular dynamics are nearly exclusively formulated as perturbative correction . Both spin-diabatic and adiabatic representations to include SOC in trajectory-based surface hopping procedures have been pioneered by Persico and Granucci. Within a perturbative description, common approximate spin-orbit Hamiltonians comprise Breit-Pauli spin-orbit operators and Douglas-Kroll transformed spin-orbit operators . Furthermore, the perturbational expansion of the Dirac equation to lowest order, the zero-order regular approximation (ZORA), is state-of-the-art and the approximate ZORA Hamiltonian was shown to provide an accurate description of spin-orbit coupling effects on properties of heavy element diatomic molecules. Van Wüllen furthermore proposed to use a model potential to construct the ZORA kinetic energy operator, circumventing the intrinsic problems of the ZORA Hamiltonian of gauge noninvariance and nonstationarity of the energy. Van Wüllen's model potential brings the advantage that the ZORA operator is then no longer dependent on the Kohn-Sham orbitals and, thus, that thereon based ionization potentials are improved, with remaining errors being comparable to original ZORA computations corrected with the electrostatic shift approximation and related to deficiencies of the exchange-correlation potential. Moreover, applying a model potential enables analytical expressions for nuclear gradients. To apply the mentioned approximate Hamiltonians not only to all-electron computations, but also to heavier-element chemistry relying on pseudopotentials, Kleinmann , Bachelet and Schlüter and Hybertsen and Louie pioneered the construction of pseudopotentials that account for spin-orbit coupling. Kleinman showed that the SOC-pseudopotentials are correctly including not only core-electron orthogonality effects, but also relativistic effects up to order α 2 with α being the fine-structure constant . Following the suggested norm-conserving construction scheme of Bachelet and Schlüter, Hartwigsen et al. furthermore constructed dual-space Gaussian pseudopotentials , featuring the general advantage of the Goedecker-Tetter-Hutter-potentials to being separable into a local and non-local part, with the integration of the latter scaling quadratically with system size. SOC-GTHpseudopotentials were parameterized with respect to two-component Dirac computations, thus including scalar-relativistic and SOC effects. They were shown to accurately reproduce bond lengths of an extended list of small molecules containing various heavy elements. In general, SOC-including PPs were also validated for a broader spectrum of molecular properties, e.g. yielding accurate predictions of spin-orbit splittings of band structures . The combination of the mentioned approximate Hamiltonians with time-dependent density functional theory (TDDFT) poses the challenge that SOC matrix elements are defined with respect to many-electron wave functions. Tavernelli et al. therefore proposed the usage of so-called auxiliary many-electron wave functions (AMEW) based on the Casida ansatz , with the auxiliary functions being constructed relying on the TDDFT response functions and thereon defined creation and annihilation operators. Corresponding implementations for the Sternheimer ansatz of LR-TDDFT within the CPMD program package were based on the Breit-Pauli Hamiltonian and conventional pseudopotentials without empirical reparameterization of the effective nuclear charges. Lowestenergy SOC matrix elements for formaldehyde were validated with respect to multi-configuration quasi-degenerate perturbation theory computations and the implementation was extended to NAMD on spin-diabatic surfaces, not requiring nuclear forces due to the perturbative ansatz, but providing SOC-corrected transition probabilities using the Landau-Zener formula. Bussy and Hutter presented an analogous implementation for core excitations enabling the efficient modeling of X-ray absorption spectra implying periodic boundary conditions , decoupling the investigated core states by considering only a small number of localized donor core orbitals. Through reduced resolution-of-the-identity basis sets located solely at the excited atom and acceleration by the auxiliary density matrix method, the authors demonstrated that the so-obtained SOC-corrected L-edge spectra are in agreement with experimental data at sub-cubic scaling of computational cost. Malis and Luber presented perturbative spin-orbit couplings for the delta self-consistent field ansatz analogously relying on the AMEW ansatz. The implementation within a mixed Gaussian and plane wave framework implying pseudopotentials was validated for formaldehyd with respect to LR-TDDFT reference values obtained with the ORCA program package . Within this context, we present an efficient implementation of a perturbative spin-orbit coupling ansatz relying on AMEW as suggested by Tavernelli et al. for the Tamm-Dancoff approximation (TDA) of linear-response time-dependent density functional perturbation theory, implemented within the mixed Gaussian and plane wave framework of the CP2K program package . The presented implementation is based on the mentioned related implementation for core excitations of Bussy and Hutter , extending the formulation to valence excitations (section 2.1) and SOC-corrected pseudopotentials of Hartwigsen et al. (section 2.2). Pseudopotential-based spin-orbit couplings are validated with respect to all-electron computations performed with the ORCA program package (section 3.3) and benchmarked for a test set of 13 organic molecules against multi-reference configuration interaction (CI) reference data (section 3.2). Computational timings and comparison with experimental spectra are outlined for the bismuth-containing CAU7 metal-organic framework (MOF) (section 3.4). |
66b51a695101a2ffa89c19fe | 1 | The perturbative SOC ansatz has been discussed in detail in Refs. . We therefore restrict ourselves to a very short introduction of the most important equations. Within the Tamm-Dancoff approximation (TDA) of time-dependent density functional theory, the excitation energy Ω N of the excited state N and the corresponding excitation amplitude X N is given as the solution of a hermitian eigenvalue problem, |
66b51a695101a2ffa89c19fe | 2 | with S representing the atomic-orbital overlap matrix. For restricted closed-shell references with α (spin up) and β (spin down) spin components being degenerate, spin adaptation by unitary transformation enables to define singlet excitation amplitudes and energies, X N sing and Ω N sing , with total spin S = 0 and spin quantum number m S = 0, which are decoupled from the corresponding triplet excitation amplitudes and energies, X N trip and Ω N trip , with total spin S = 1 and m S = 0. In the following, spin-adapted matrices will be indicated with lower indices M Sm S , labeling total spin S and spin angular momentum m S . The eigenvalue equations for singlet and triplet spin are given after spin summation within the Sternheimer formalism as |
66b51a695101a2ffa89c19fe | 3 | with a S = 2 for singlet, and a S = 0 for triplet spin. The Kohn-Sham matrix contributions F and kernel contributions K depend on the chosen functional and comprise one-electron, Coulomb and exchange-correlation (XC) contributions. The projection operator onto the virtual orbital space Q ensures orthonormality of excitation amplitudes X N with the ground-state MO coefficients C. The global parameter a EX scales the amount of exact exchange and the two-electron repulsion integrals are defined in Mulliken notation. |
66b51a695101a2ffa89c19fe | 4 | The remaining triplet states with spin angular momentum m S = 1 and m S = -1 require to take into account spin-flipped excitations, which are not accessible via the collinear TDA eigenvalue problem given in Eq. ( ). Tavernelli et al. therefore suggested to rely on auxiliary many-electron wave functions (AMEW) , which define singlet and triplet states as singly excited determinants, |
66b51a695101a2ffa89c19fe | 5 | with X N sing and X N trip referring to the spin-adapted singlet and triplet excitation amplitudes and a, b, c, d, . . . indicating virtual orbitals. |Ψ 0 denotes the KS reference Slater determinant built from the ground-state KS orbitals φ iσ . As noted by Marian , Eqs. ( )-( ) can be obtained by choosing the AMEW with maximum spin quantum number and generating the remaining states by applying the Wigner-Eckhart theorem. |
66b51a695101a2ffa89c19fe | 6 | While orthonormality of excitation amplitudes X N Sm S of same spin as well as orthonormality of both amplitudes with the ground-state MO coefficients C is implied, an orthonormality for amplitudes of different spin, i.e. between X N sing and X N trip , is not given due to the fact that both eigenvalue equations and corresponding constraints are decoupled (Eq. ( )). For mixed matrix elements, as given in Eq. ( ), the Slater-Condon rules therefore do not apply and Eq. ( ) has to be solved relying on the more general Löwdin rules defining the required matrix elements as outlined in Refs. . The one-electron SOC-Hamiltonian ĥSO within the zeroth-order regular approximation (ZORA) is defined as |
66b51a695101a2ffa89c19fe | 7 | with the potential V being approximated as Van Wüllen's model potential. c is the speed of light and p = -i∇ the momentum operator matrix. σ P represents the vector of Pauli spin matrices. Final resulting couplings between singlet and triplet states of different spin quantum numbers are given in the supplementary information. Alternatively, if restricting the description to valence electrons only, a SOC-corrected pseudopotential, as parameterized by Hartwigsen et al. , can be used, |
66b51a695101a2ffa89c19fe | 8 | and L representing the angular momentum matrix and ∆V SO l (r, r ) representing the non-local SOC correction pseudopotential contribution to the remaining scalar relativistic local and non-local terms. Explicit formula are given in Ref. . The reduction to one-electron matrix elements over molecular orbitals relying on the Löwdin formula and integration of spin can be performed as for the all-electron Hamiltonian. Diagonalizing the resulting matrix of spin-orbit matrix elements (SOME) H NM Sm S ,S m S of dimension Ñ × Ñ with Ñ = (1 + N + 3N) gives the SOC-corrected excitation energies Ω Ñ as well as the perturbed eigenfunctions Ψ Ñ . Resulting SOC-corrected eigenvectors can be used to compute the corresponding oscillator strengths f len Ñ within the length or velocity representation, with rζ indicating the dipole operator and pζ the momentum operator. Both representations are equal for exact wave functions not including a non-local potential at the complete basis-set limit or for approximate wave functions if fulfilling the Thomas-Reiche-Kuhn sum rule at the complete basis-set limit. Explicit working equations for the AMEW ansatz are given in the supplementary information. |
66b51a695101a2ffa89c19fe | 9 | To validate the implementation within the CP2K program package, splittings and oscillator strengths relying on the all-electron ZORA-Hamiltonian of Eq. ( ) were compared with the analogous implementation in ORCA featuring perturbative SOC relying on a Breit-Pauli-Hamiltonian . All calculations were performed using the Perdew-Burke-Ernzerhof with 25% Hartree-Fock exchange (PBE0) functional and TZP-ZORA basis sets . The first 10 excited states of the homologue series of H |
66b51a695101a2ffa89c19fe | 10 | Spin-orbit splittings and spin-orbit coupling matrix elements (SOMEs) were furthermore validated for a benchmark set by Dinkelbach et al. The test set originally comprises 14 organic molecules. For two of the therein included structures, namely furan and thiophen, Dinkelbach et al. explicitly added diffuse basis functions centered at a dummy atom within the five-membered aromatic ring. For convenience and to enable an unbiased validation of all test structures, furan and thiophen were therefore excluded in our benchmark studies. The remaining 12 test molecules are visualized in Fig. , featuring C, O, N, S, and H atoms. |
66b51a695101a2ffa89c19fe | 11 | The performed TDA computations are compared with TDA and DFT/MRCI reference results of Ref. , with the latter representing an approach combining KS orbitals with restricted multi-reference configuration interaction (MRCI) expansions. For TDA spin-orbit couplings, Dinkelbach et al. were also relying on the AMEW ansatz and thus their TDA reference results are based on comparable wave functions. In contrast to the current implementation, the reference results are however based on an atomic meanfield approximation of the Breit-Pauli Hamiltonian, thus including, next to spin-same also spin-other-orbit interactions. |
66b51a695101a2ffa89c19fe | 12 | Triple-ζ basis functions with one polarization function from the def2 basis set family of the Ahlrichs group were used for all atoms of the first and second row of the periodic table (TZVP), while an additional polarization function was added for sulfur (TZVPP). In addition, augmented basis sets were used for formaldehyde, thioformaldehyde and nitromethane to account for low-lying Rydberg states. Since CP2K is not featuring an assigment of excited states regarding symmetry, the assignment of excited states was based on a comparison of CP2K results with analogous Turbomole computations .In general, it thus has to be highlighted that TDA and DFT/MRCI results were compared and corresponding errors were computed based on the symmetry assignment and not on the energetic ordering. Only for dithiin, nitrobenzene, nitromethane and o-benzyne, symmetry assignment was in full agreement with an ordering of excited states based on the excitation energy. For bithiophene, dithiosuccinimide and isoalloxaline, the energetic ordering of two states and for quinoxaline the ordering of three pairs of states differs from the reference assignment, increasing the listed error in the excitation energy. Resulting errors in the excitation energies amount up to 0.5 eV for both 1 1 B 2 and 2 1 A 2 of dithiosuccinimide and 1 1 A 2 and 2 1 A 2 for quinoxaline. Comparable deviations are discussed for full TDDFT not relying on the TDA in Ref. , suggesting the multireference character of these states to be significant. Further details on the assignment are given in the supplementary information. |
66b51a695101a2ffa89c19fe | 13 | Overall errors of excitation energies and SOMEs with respect to the reference TDA and DFT/MRCI results are summarized in Table . Corresponding correlation plots for excitation energies and spin-orbit coupling matrix elements are given in Figures and. In sum, total RMSDs with respect to the reference TDA@PBE0 results are of minor 0.10 eV for excitation energies and of 3.7 cm -1 for spin-orbit coupling elements. Deviations with respect to DFT/MRCI references amount up to total RMSDs of 0.28 eV for excitation energies and 5.9 cm -1 for spin-orbit coupling elements. Due to the neglect of spin-other-orbit interactions in the Hamiltonian and differences in basis sets and setup, relying on the Gaussian and augmented plane wave implementation with all-electron potentials in comparison to all-electron implementations for both reference data sets, the TDA@PBE0 results of this work are considered less accurate by construction in comparison to the TDA reference. The nevertheless good agreement with the methodologically most accurate DFT/MRCI results must therefore rely on fortuitious error cancellation. In case of the SOME, two points clearly deviate from Fig. (right) at 137.2 cm -1 and 83.6 cm -1 . These belong to 1 3 A 2 |H SO z |1 1 A 1 and 1 3 B 2 |H SO x |2 1 A 1 from thioformaldehyde. For 1 3 A 2 |H SO z |1 1 A 1 , the deviation between MRDFT and the reference TDA-TDDFT calculation is with 37.1 cm -1 much larger than our calculation with only 8.4 cm -1 . This could be due to a compensation of errors between the fully Gaussian basis set used in turbomole and the Gaussian and augmented plane waves approach used in CP2K. For 1 3 |
66b51a695101a2ffa89c19fe | 14 | Table . Root-mean-square deviations (RMSDs) in excitation energies [eV] and spin-orbit coupling matrix elements (SOMEs) [cm -1 ] of TDA@PBE0 results with respect to reference TDA@PBE0 and DFT/MRCI results taken from Ref. . For comparison, RMSDs comparing the reference TDA@PBE0 computations with the reference DFT/MRCI results are listed in the last column. |
66b51a695101a2ffa89c19fe | 15 | For computations on systems involving heavier elements, core electrons are treated by adding the SOC-including pseudopotential of Eq. ( ) and thus by treating solely valence electrons explicitly. An extensive benchmark of SOC-including pseudopotentials is beyond the scope of this work, however, the implementation was validated for the subset of small organic molecules discussed in section 3.1. To enable a straight-forward comparison, computations were performed using aug-QZV2P-GTH basis sets and we therefore restricted the validation to molecules containing solely atoms up to the third row of the periodic table. Overall RMSDs for the comparison of pseudopotential computations against the all-electron reference, given in Table , are smaller than 0.28 eV, respectively. RMSDs in SOMEs are in the range of 10.4-114.5 cm -1 . |
66b51a695101a2ffa89c19fe | 16 | Timings are presented for a metal-organic framework, dubbed CAU7-TABA, containing 184 atoms in the unit cell corresponding to 656 valence electrons and 2600 basis functions. Computations were performed on a 512 Intel Xeon Gold 6426Y processor using in total 512 cores. Using a PBE/TZVP-MOLOPT-GTH basis set and the mentioned SOC-corrected GTH pseudopotentials and computing 200 excited states, total computation time amounted up to 7h 17 min 23.1 s. Computation of the splittings and spin-orbit coupling matrix elements took only 14 min 58.1 s of the total computation time. |
66b51a695101a2ffa89c19fe | 17 | Perturbative spin-orbit couplings were presented for the Tamm-Dancoff approximation of linear-response time-dependent density functional theory, relying on the AMEW ansatz and a one-electron ZORA Hamiltonian as well as SOC-corrected pseudopotentials. All-electron spin-orbit splittings and couplings are in agreement with DFT/MRCI reference results with errors in excitation energies and spin-orbit matrix elements being below 0.2 eV and 5.8 cm -1 , respectively. Relying on pseudopotentials, root-mean-square deviations increase by an order of magnitude. Computation of the perturbative correction represents less than 3% of the overall computation time. The ansatz could thus be beneficial for non-adiabatic molecular dynamics, in particular the description of intersystem crossing. |
67b1e78a81d2151a02b673b7 | 0 | Metal-halide perovskites (MHPs) have emerged as highly promising materials due to their exceptional properties and versatile applications in photovoltaics, optoelectronics, and photocatalysis. These compounds possess exceptional optical and electronic properties, including a high absorption coefficient, a direct and tunable bandgap, long carrier diffusion lengths, high charge mobility, and low recombination losses. The general crystal structure of MHPs is ABX3, where A is typically an organic or inorganic cation (Methylammonium, Cs or Rb), B is a metal cation (Pb or Sn), and X is a halide anion (Br, I, or Cl). The structure consists of BX₆ octahedra, with the B cation at the center and the A cation occupying a 12-fold coordinated site. Depending on composition and conditions, perovskites can crystallize in cubic, orthorhombic, or tetragonal systems. Among other MHPs, cesium lead bromide (CsPbBr₃) perovskite has drawn considerable interest due to its superior thermal and structural stability compared to its organic and inorganic counterparts. CsPbBr₃ (CPB) can exist as bulk crystals or nanocrystals (NCs), each exhibiting distinct properties suited for different applications. Single crystals of CPB offer excellent charge transport properties, low defect densities, and high carrier mobility, making them ideal for optoelectronic devices such as |
67b1e78a81d2151a02b673b7 | 1 | photodetectors and solar cells. However, their large size limits their processability and surface area, which is critical for applications like photocatalysis and light emission. In contrast, CPB NCs exhibit outstanding optoelectronic properties, such as high photoluminescence quantum yield, high charge carrier mobility, and high absorption coefficient in the visible light range. These materials are widely used in photocatalysis, fluorescence sensing, and optoelectronic devices. Despite these promising features, one of the most critical challenges for CPB NCs remains their stability. The ionic crystal structure makes them less resistant to chemical and physical vulnerabilities such as heat, polar solvents, light, and oxygen. CPB NCs are highly sensitive to polar solvents and harsh reaction environments, which can lead to rapid degradation. Researchers have explored various strategies, including surface passivation, encapsulation, and interface engineering, to address the instability of perovskite-based devices and extend their operational lifespan. In this context, metal-organic frameworks (MOFs) have emerged as a promising approach to significantly enhancing CPB stability by encapsulation. Structurally, MOFs consist of single or mixed metal ions coordinated with organic linkers (ligands) to form a porous framework, enabling applications in gas storage, supercapacitors, and photocatalytic dye degradation. The MOF materials pose high crystallinity, large surface area, excellent porosity, and remarkable environmental stability. Thus, these materials can effectively shield CPB from external environmental factors by providing a protective barrier, thereby reducing degradation and improving long-term performance. Furthermore, MOFs are among the most effective catalysts for photocatalysis due to their large surface area and strong lightharvesting ability. These properties significantly enhance the performance of CPB-MOF-based nanocomposites, particularly in photocatalytic applications. Typically, MOFs synthesized under mild conditions are preferred for composite formation with CPB NCs to prevent their degradation under harsh conditions. Zeolitic Imidazole |
67b1e78a81d2151a02b673b7 | 2 | Framework-8 (ZIF-8) is particularly well-suited for this role due to its adaptable synthesis routes tailored to specific requirements. It can be synthesized under mild conditions, such as room temperature, and exhibits exceptional properties like high porosity and a large surface area. Moreover, ZIF-8 is compatible with nonpolar solvents, making it well-suited for integrating with CPB without compromising its stability. Another interesting aspect of previous studies is that most CPB/ZIF-8 nanocomposites have been synthesized using the ligand-assisted reprecipitation (LARP) method. |
67b1e78a81d2151a02b673b7 | 3 | Notably, most studies focus on encapsulating CPB NCs with MOFs to enhance their stability and performance. While encapsulation has shown promise in enhancing CPB's environmental stability, it faces challenges such as incomplete pore filling, partial encapsulation, and limited control over nanocrystal size. Additionally, synthesizing MOFencapsulated CPB is typically complex and involves multiple steps. Furthermore, complete encapsulation can hinder efficient charge transfer between CPB and its surroundings, potentially compromising performance, including low quantum yield, and compromising the activity of this material as a photocatalyst. Decorating CPB NCs onto MOFs presents a highly effective approach for photocatalytic applications, offering several key advantages. This method preserves the high quantum yield of CPB NCs without compromising their optical properties while simultaneously enhancing the optical activity of both components. The exposed NCs on the MOF surface facilitate superior light absorption and charge separation, significantly boosting photocatalytic performance. Additionally, this structured decoration enhances charge transfer between CPB NCs and MOFs, a crucial factor for efficient photocatalysis. The ability to precisely control NC size allows for fine-tuning of absorption bands, optimizing the material for specific light wavelengths. Furthermore, the decoration technique ensures uniform NC distribution, maintaining consistent optical properties throughout the composite. Beyond photocatalysis, this approach holds great potential for applications such as sensing, where the high quantum yield and improved charge transfer efficiency of CPB NCs can be fully leveraged without the constraints of full encapsulation. This versatility makes the decoration method a powerful strategy for designing advanced materials for photocatalysis, optoelectronics, and other highperformance applications. |
67b1e78a81d2151a02b673b7 | 4 | Despite the significant advantages of uniformly decorated CPB NCs over ZIF-8, achieving a stable and well-distributed nanocomposite remains challenging. Our research focuses on elucidating three critical aspects: the decoration mechanism, the nature of charge transfer, and the proof of the photocatalytic mechanism. A comprehensive understanding of these processes is essential for optimizing material performance and advancing CPB/ZIF-8-based photocatalytic applications. The precise mechanism of CPB NCs integration with ZIF-8 remains unclear. While ligand exchange is a possible pathway, alternative processes may also contribute to the decoration process. For instance, the confined growth of ZIF-8 in mesoporous structures suggests the possibility of CPB NCs nucleating and growing within ZIF-8 pores. Gaining insight into this mechanism is crucial for ensuring uniform decoration and composite stability. Precise control over ligand exchange and surface interactions is necessary to facilitate proper integration, minimize defects, and enhance charge transfer efficiency. A deeper understanding of these interactions will not only streamline synthesis but also enable the fine-tuning of photocatalytic performance, paving the way for scalable production and commercialization. Additionally, the charge transfer dynamics between CPB NCs and ZIF-8 remain largely unexplored, yet they play a crucial role in determining photocatalytic efficiency. Here, we present a straightforward and effective synthesis method for preparing a uniformly distributed CPB/ZIF-8 nanocomposite. Fourier-transform infrared (FTIR) spectroscopy analysis confirmed that ligand exchange is the primary driving force behind composite formation, where the native ligands of CPB NCs were replaced by the imidazole ligands of ZIF-8. This provided valuable insights into the chemical interactions within the composite. |
67b1e78a81d2151a02b673b7 | 5 | The unshifted absorbance spectra of CPB NCs in the composite indicated the preservation of their structural integrity. To investigate charge transfer within the composite, photoluminescence (PL) studies were conducted, revealing an impressive 92% PL quenching, signifying highly efficient charge transfer between CPB NCs and ZIF-8. Additionally, electron paramagnetic resonance (EPR) analysis provided deeper insights into charge dynamics, demonstrating enhanced radical generation, particularly of hydroxyl radicals (•OH). This finding highlights not only efficient charge transfer but also superior charge separation, |
67b1e78a81d2151a02b673b7 | 6 | reinforcing the composite's potential for advanced environmental remediation. The photocatalytic performance of the composite was evaluated using dye degradation as a model reaction, both to assess its real-world applicability and to further understand the charge transfer process. The CPB/ZIF-8 composite exhibited significantly enhanced dye degradation efficiency compared to its pristine components, validating its improved photocatalytic activity. |
67b1e78a81d2151a02b673b7 | 7 | ZIF-8 was synthesized based on the previous reports. Specifically, 297 mg (1 mM) of zinc nitrate (Zn (NO3)2•6H2O) and 656 mg (8 mM) of 2-methylimidazole were dissolved separately in 12 mL of methanol. The 2-methylimidazole solution was slowly added to the zinc nitrate solution and stirred for 6 h at room temperature (RT). The resulting white precipitate was collected by centrifugation (6000 rpm,10 min) and washed with methanol (3×30 mL) to remove excess ligands. The white residue was dried overnight in a vacuum oven at 70 ºC to obtain ZIF-8. |
67b1e78a81d2151a02b673b7 | 8 | Preparation of Cs-oleate: The CsPbBr3 NCs were synthesized based on the previously reported methods. Generally, cesium oleate solution was prepared by dissolving 40. Synthesis of CsPbBr3 NCs: Once PbBr2 was fully dissolved, the N₂ atmosphere was temporarily removed, and 0.4 mL of the previously prepared cesium oleate solution was swiftly injected into the lead precursor solution at 170 °C. Within 5 seconds, the reaction mixture was quenched by immersing the flask in an ice-water bath to rapidly reduce the temperature to below RT. The solution was then equilibrated to RT, ensuring the transition to a solution phase. |
67b1e78a81d2151a02b673b7 | 9 | The CsPbBr₃/ZIF-8 (CPB/ZIF-8) nanocomposite was synthesized using a straightforward mixing process. Initially, 30 mg of ZIF-8 powder was dispersed in 2 mL of hexane and sonicated for 5 minutes to achieve uniform dispersion. Subsequently, 3 mL of CsPbBr₃ (CPB) solution in hexane was introduced into the ZIF-8 dispersion. The mixture was again sonicated for 5 minutes to ensure uniform mixing, followed by stirring under a vacuum at 50°C overnight to facilitate the efficient incorporation of CPB NCs into the ZIF-8 framework. The precipitate formed was collected by centrifugation at 9,000 rpm for 10 minutes and washed with hexane (2× 10 mL) to remove the excess NCs. Finally, the purified CPB/ZIF-8 nanocomposite was dried under vacuum at 50°C for 6 h. This method yielded a highly uniform CPB/ZIF-8 nanocomposite, ensuring effective interaction between the CPB NCs and the ZIF-8 framework. |
67b1e78a81d2151a02b673b7 | 10 | The X-ray diffraction (XRD) pattern was recorded in the 2θ range 5-50º using a Rigaku SmartLab diffractometer with a Cu anode (Cu Kα radiation with 1.5418 Å wavelength) operated at 40 kV and 120 mA. The morphology of the samples was observed via Highresolution scanning electron microscopy (HR-SEM) using a Tescan MAIA3 instrument equipped with an energy dispersive spectrometer (EDS) detector and Thermo Fisher Scientific™ Talos™ F200X transmission electron microscope (TEM)). The FTIR-6800, Jasco, Japan, was used to perform Fourier transform infrared spectrometer (FTIR) measurements in the spectral region of 3500 to 400 cm -1 . |
67b1e78a81d2151a02b673b7 | 11 | Electron paramagnetic resonance (EPR) spectra were recorded on a Bruker ELEXSYS 500 Xband spectrometer equipped with a Bruker ER4102ST resonator in a Wilmad flat cell for aqueous solutions (WG-808-Q) at RT. Experimental conditions were 512 points, with a microwave power of 20 mW, 0.1 mT modulation amplitude, and 100 kHz modulation frequency. The sweep range was 200 mT. |
67b1e78a81d2151a02b673b7 | 12 | Photocatalytic dye degradation experiments were conducted using a Xenon lamp with a power output and light intensity of 485 mW/cm², equipped with a UV cut-off filter. A solvent mixture of 25 mL consisting of 10% (v/v) ethanol in toluene was used for all experiments. The concentrations of the catalyst and dye were maintained at 400 mg/L and 16 mg/L, respectively, across all trials to ensure consistency. Prior to degradation, the catalyst and dye were dispersed via sonication for 5 minutes to ensure homogeneity. All measurements were performed at ambient temperature under continuous stirring to maintain uniform reaction conditions. To address the high pH dependency, we stirred the BCG solution in the dark for 30 minutes following the addition of the catalyst to allow pH stabilization.. The micrograph of the CPB/ZIF-8 nanocomposite reveal the distribution of CPB particles across the ZIF-8 surface (Figure ). Further TEM analysis (Figure ) confirms this result, |
67b1e78a81d2151a02b673b7 | 13 | showing dark dot-like structures identified as CPB NCs decorating the ZIF-8 surface. The lattice spacing of CPB NCs (Figure ), measured at 0.42 nm, corresponds to the (110) planes of the orthorhombic phase of the CPB crystal structure, providing additional confirmation of the nanocomposite formation. Notably, this uniform distribution was not a straightforward technique. The primary challenge lies in the instability of CPB NCs in polar solvents, coupled with the poor dispersibility of ZIF-Achieving a uniform dispersion of CPB NCs within the ZIF-8 framework presents a significant challenge due to their strong tendency to aggregate, which can compromise structural homogeneity and alter optical properties. The prolonged stirring and slight temperature increase improved the distribution of CPB NCs over the ZIF-8 surface in several ways. First, prolonged stirring ensures better dispersion of precursors, enhancing the interaction between components like CPB NCs and ZIF-8. Additionally, thorough mixing during extended stirring forms a more stable nanocomposite. Secondly, the elevated temperature increases the kinetic energy of the NCs, causing them to move more rapidly and vigorously. The higher molecular speed results in more frequent collisions between reactant particles, enhancing the likelihood of successful interactions. Overall, prolonged stirring and a slight temperature increase synergistically promote a more uniform and stable distribution of CPB NCs over the ZIF-8 surface, enhancing the nanocomposite's quality. |
67b1e78a81d2151a02b673b7 | 14 | To better understand the nanocomposite formation, we analyzed the sizes of the CPB NCs on the ZIF-8 and in the supernatant after mixing. The supernatant was collected via centrifugation following the final stage of nanocomposite formation between CPB NCs and ZIF-8. TEM analysis showed that the average size of CPB NCs attached to ZIF-8 was 15.54 ± 3.5 nm (Figure ). The size distribution graphs exhibit symmetry with peaks centered around the mean, indicating the uniformity in the sizes of the CPB NCs on the ZIF-8. Notably, the average CPB NC size in the nanocomposite remained consistent with that of the pristine solution, which measured 15.55 ± 2.6 nm. |
67b1e78a81d2151a02b673b7 | 15 | The average size of CPB NCs in the supernatant (Figure ) decreased to 12.46 ± 5.7 nm, indicating that larger CPB particles preferentially adhered to the ZIF-8 surface during nanocomposite synthesis. This phenomenon can be attributed to the tendency of larger CPB NCs to detach from their ligands and adhere to template surfaces such as MOFs. The primary factor driving this "ligand-exchange" behavior is surface energy; larger NCs possess a greater surface area, resulting in higher surface energy, which they reduce by detaching from ligands and attaching to more stable surfaces. Additionally, as NCs grow, their surface-to-volume ratio decreases, resulting in fewer ligand binding sites relative to their size and weaker overall ligand attachment. Electrostatic interactions also play a role, as larger CPB NCs typically have a more negative surface charge, creating an initial attraction to the positively charged surfaces or metal ions in ZIF. This is followed by ligand exchange between the ligand molecules in ZIF and the positively charged components of the CPB NCs, resulting in chemical interaction. These factors collectively drive the attachment of larger NCs to MOFs with almost uniform sizes. |
67b1e78a81d2151a02b673b7 | 16 | The area of CPB coverage over ZIF-8 was quantified using TEM imaging (Figure ) and ImageJ software. The results indicate that approximately 24% of ZIF-8 was covered by CPB NCs. The micrographs revealed a well-spaced distribution of CPB NCs with uniform sizes, a characteristic not reported in previous studies, which typically showed either full coverage or uneven distributions of CPB NCs with varying sizes. TEM analysis indicated that slightly larger CPB nanocrystals adhere to the ZIF-8 framework, exhibiting a uniform size distribution. |
67b1e78a81d2151a02b673b7 | 17 | These larger crystals are stabilized by the framework due to surface energy, ligand exchange, and electrostatic interactions. Structural analysis XRD XRD analysis was performed on ZIF-8, pristine CPB, and CPB/ZIF-8 to examine the formation of nanocomposites and their structural properties, as illustrated in (Figure ). The diffraction pattern of CPB and ZIF-8 (Figure , red and blue line, respectively) exhibits well-defined peaks corresponding with the previously reported crystalline structure (JCPDS -01-072-7929 and JCPDS -00-062-1030, respectively). Notably, due to the random orientation of CPB NCs, the typical shoulder peaks at 15.2º, 21.5º, and 30.5º are indistinguishable. Interestingly, the CPB/ZIF-8 nanocomposite exhibits only the characteristic ZIF-8 and CPB characteristic peaks for ZIF-8 and CPB. Indeed, The sharp, well-defined peaks of ZIF-8 remain intact without any noticeable broadening, confirming that its crystalline structure was preserved in the nanocomposite. The XRD results confirm the successful formation of the CPB/ZIF-8 nanocomposite, with both materials retaining their characteristic crystalline structures as evidenced by the sharp ZIF-8 peaks and distinct CPB reflections. exhibited a prominent peak at 420 cm⁻¹, corresponding to the stretching vibration of the Zn-N bond between Zn²⁺ and the 2-methylimidazole ligand. Pristine CPB displayed characteristic peaks at 2848 cm⁻¹ and 2917 cm⁻¹, representing its organic ligands' C-H bending and stretching vibrations. In the FTIR spectrum of the CPB/ZIF-8 nanocomposite, peaks corresponding to both ZIF-8 and CPB were observed. The ZIF-8 peaks exhibited significantly higher intensity than those of CPB, indicating a higher concentration of ZIF-8 per unit volume. Interestingly, the CPB ligand peaks in the 2850-2950 cm⁻¹ range showed a significant reduction in intensity, supporting ligand exchange alongside surface charge interactions. This reduction indicates interactions between Pb²⁺ ions from the CPB NCs and the nitrogen atoms of the imidazole ligands in ZIF-8, enabling both chemical and physical adsorption. |
67b1e78a81d2151a02b673b7 | 18 | The optical properties of the pristine materials and nanocomposites were analyzed using UV-Vis absorption and PL spectroscopy. The absorption spectrum of ZIF-8 (Figure ), measured in methanol, shows a distinct maximum at 210 nm. Due to the poor stability of CPB NCs and CPB/ZIF-8 nanocomposites in polar solvents, hexane was used as the dispersion medium for these measurements. However, as hexane is UV-active, absorption below 300 nm could not be recorded. The pristine CPB NCs and CPB/ZIF-8 exhibit absorption edges at 512 nm and 511 nm, respectively (Figure ). The negligible shift in the absorption edge between pristine CPB and the nanocomposite confirms that the optical activity of the CPB NCs remains intact within the CPB/ZIF-8. |
67b1e78a81d2151a02b673b7 | 19 | The Tauc equation was used to calculate the materials' optical band gap (Eg), and the corresponding graphs are provided in the supplementary information (Figure ). The band gap (Eg) values of ZIF-8, CPB NCs, and CPB/ZIF-8 were determined to be 5.34 eV, 2.32 eV, and 2.32 eV, respectively. The optical band gap analysis indicates no significant change between the CPB NCs and the CPB/ZIF-8 composite, confirming that the CPB NCs contribute significantly to the composite's optical activity. |
67b1e78a81d2151a02b673b7 | 20 | Figure presents the PL spectra of CPB NCs and CPB/ZIF-8 composites. The CPB NCs exhibit an intense green fluorescence peak at 513 nm, while the PL emission peak of CPB/ZIF-8 is observed at 517 nm. The slight red shift observed in the PL of the CPB/ZIF-8 nanocomposite can be attributed to two primary factors: a) Attachment of larger CPB NCs: |
67b1e78a81d2151a02b673b7 | 21 | The interaction between the CPB NCs and the ZIF-8 framework can alter the local dielectric environment around the nanocrystals, further contributing to the shift. Additionally, the CPB/ZIF-8 nanocomposite showed a significant reduction in PL intensity, quenching by 92% compared to pristine CPB NCs. This quenching effect can be attributed to charge transfer from the CPB to the ZIF-8 framework, as depicted in Scheme 3. The primary cause of this charge transfer is likely the interaction between Pb²⁺ ions in CPB and nitrogen atoms in the imidazole ligands of ZIF-8. . This observation aligns with the earlier FTIR results, which indicated ligand exchange. |
67b1e78a81d2151a02b673b7 | 22 | An EPR analysis was conducted to analyze the radical formation and stability of the CPB NCs and CPB/ZIF-8 composite. This experiment used the 5-tert-butoxycarbonyl-5-methyl-1pyrroline-n-oxide (BMPO) as the spin trap. Interestingly, the CPB and CPB/ZIF-8 composite were photoactive and showed strong EPR signals (Figure ). This indicates that both materials can generate charge carriers (electrons and holes) when exposed to light. The CPB/ZIF-8 composite exhibited a stronger EPR signal than pristine CPB NCs. This enhancement suggests a higher concentration of unpaired electrons or radicals in the composite system. The increased radical formation in the composite can be attributed to several factors. One is improved charge separation; the ZIF-8 framework may facilitate better separation of photogenerated electrons and holes, reducing their recombination rate and allowing more charges to participate in radical formation. That assumption is also supported by the PL results presented above. Another factor might be the enhanced stability; the higher stability of CPB NCs within the ZIF-8 framework allows for prolonged photoactivity and sustained radical formation. The EPR signals detected in both CPB and CPB/ZIF-8 composite exhibited a distinct quartet peak, indicating the presence of hydroxyl radicals (•OH). To confirm the formation of the •OH, DMSO was added to the samples as a quencher, and EPR signals were analyzed. After adding the DMSO, the EPR signals were quenched, confirming the formation of •OH during the process. The EPR measurements are carried out in the dark and under illumination (Tungsten lamp) conditions in the 10% (v/v) EtOH/Toluene solution (Figure ). The ZIF-8 was found to be EPR inactive, and it is due to the presence of solely Zn 2+ ions in its crystal structure. The EPR analysis reveals that the CPB/ZIF-8 composite exhibits enhanced stability and superior radical formation capabilities, particularly in generating •OH. This stability and increased radical formation efficiency can make the CPB/ZIF-8 composite a promising material for various photocatalytic applications, potentially outperforming pristine CPB NCs in effectiveness and durability. (10% (v/v) EtOH/toluene) conditions used for EPR analysis were also maintained for the dye degradation studies. The characteristic peaks at 421 nm for MO and 610 nm for BCG were used to monitor dye degradation. BCG also showed an additional peak at 412 nm, attributed to its monoanionic form and the pH dependency of the solution. . Both dyes were stable under illumination and dark conditions without a catalyst (Figures and). All dye degradation results indicated that the CPB/ZIF-8 nanocomposite exhibited superior photocatalytic activity compared to CPB NCs and ZIF-8. Specifically, CPB/ZIF-8 achieved 1.48 and 1.75 times higher degradation efficiency than CPB NCs within 30 minutes for MO and BCG dye, respectively (Figure and S17a). The CPB/ZIF-8 composites demonstrated a significant 1.5-times improvement in activity compared to previously reported CPB/MOF composites in the degradation of MO. The degradation kinetics studies also revealed that both CPB/ZIF-8 and CPB NCs have good linear coefficients, indicating that they follow pseudo-first-order kinetics for both dyes (Figures and). The calculated parameters, such as the apparent rate constant, quantum efficiency (QE), and electrical energy consumption (EEC), were superior in the CPB/ZIF-8 nanocomposites. All the results regarding the photocatalysis are summarized in Table , with a detailed discussion included in the supplementary information. The superior performance of the CPB/ZIF-8 nanocomposite can be attributed to the synergistic effect between CPB and ZIF-8, which enhances stability, increases •OH formation, and improves overall catalytic efficiency. This makes the CPB/ZIF-8 nanocomposite a promising material for environmental remediation applications, particularly in degrading organic dye pollutants. |
67b1e78a81d2151a02b673b7 | 23 | This study highlights the successful development of a well-decorated CPB/ZIF-8 nanocomposite, demonstrating significant improvements in photocatalytic performance. We achieved a uniform and well-distributed composite with enhanced functionality by employing the hot-injection method for CPB NC synthesis and integrating them with ZIF-8 through an optimized mixing technique. Electron microscopy (EM) analysis revealed a remarkable distribution of larger CPB NCs across the ZIF-8 framework, exhibiting a uniform and orderly coverage that suggests a highly controlled attachment mechanism. This well-defined assembly is likely driven by surface charge interactions, emphasizing the nanoscale precision of the integration process. Fourier-transform infrared spectroscopy (FTIR) analysis provided key insights into the attachment mechanism, showing a notable reduction in CPB ligand peak intensity. This spectral shift confirms ligand exchange while highlighting the crucial role of surface charge interactions in enabling the incorporation of CPB NCs into the ZIF-8 structure. |
67b1e78a81d2151a02b673b7 | 24 | These observations emphasize the critical interaction of chemical and physical forces at the nanoscale interface. The unshifted absorption spectra also confirmed that the CPB NCs remained structurally intact within the composite. The composite exhibited a remarkable 92% PL quenching, indicating efficient charge separation. FPL spectra confirm that the optical activity of CPB NCs is preserved while benefiting from improved charge separation and reduced electron-hole recombination. Notably, this study is the first to report the photocatalytic dye degradation performance of CPB/ZIF-8, demonstrating a significant enhancement under visible light irradiation, with degradation rates 1.48 and 1.75 times higher for MO and BCG, respectively, compared to pristine CPB NCs. The EPR studies revealed superior radical generation capabilities, particularly for hydroxyl radicals, highlighting its potential for advanced environmental remediation applications. These findings underscore the synergistic interaction between CPB NCs and ZIF-8, making the CPB/ZIF-8 nanocomposite a promising material for photocatalysis and other optoelectronic applications. |
67339b2df9980725cff94c52 | 0 | When we think about the future chemistry laboratory, we often envision a fully automated system: scientists enter a molecule they want to create, advanced software suggests a synthetic route, which is then executed by a robot. This is followed by a purification step to prepare the sample for analysis by a range of analytical instruments. Finally, an automated process determines the structure of the molecule based on the spectral data collected. Achieving this vision requires advancements in multiple areas, from route predictions to synthesis and purification automation. A critical component of this pipeline is the automated interpretation of spectroscopic data for molecular structure elucidation. In this work, we describe our contribution to solve automatic structure elucidation by proposing a flexible model that can translate spectroscopic data directly into molecules, while addressing several limitations of current Computer-Assisted Structure Elucidation (CASE) programs. CASE programs use Nuclear Magnetic Resonance (NMR) and/or Infrared (IR) data to elucidate molecular structures. H and C NMR are the most widely used spectroscopic techniques for characterizing molecules and provide detailed insights into hydrogen and carbon atoms. Furthermore, 2D techniques such as COSY (Correlation Spectroscopy) and HSQC (Heteronuclear Single Quantum Coherence) provide connectivity information. Mass spectrometry (MS) offers molecular weight and formula information, crucial for confirming molecular structures. IR spectroscopy is cost-effective and non-destructive allowing for quick identification of functional groups. The fingerprint region (400-1500 cm -1 ) of IR spectra contains complex, molecule-specific absorption patterns that are challenging to interpret using traditional methods. This region holds detailed molecular information often underutilized in traditional analysis due to its complexity. However, modern machine learning algorithms show promise in extracting and interpreting this rich structural information from the fingerprint region. Together, NMR, IR, and MS provide complementary information essential for structure elucidation. |
67339b2df9980725cff94c52 | 1 | Classic CASE software packages, such as ACD/Structure Elucidator (ACD Labs) and Mnova Structure Elucidator (Mestrelab Research), have been developed to aid in this process. Typically, the user inputs processed spectral data, including peak assignments and correlations, and the software generates candidate structures that fit the data. These programs often employ spectral comparisons as part of their structure elucidation process, comparing input spectra against databases of known compounds or simulated spectra to build up structural fragments. However, these programs often require substantial human input, especially in isolating relevant peaks within NMR spectra. Additionally, they rely on databases of chemical structures and spectra, which can fail when experimental conditions alter spectra or when the database does not cover the required chemical space. Spectral comparison methods, such as the Goodman DP4 method and its variants, along with newer neural network-based and grid-based approaches, have also emerged as valuable tools for verifying molecular structures among different options. While these methods are not standalone structure elucidation techniques, they play a crucial role in structure verification and can be integrated into broader elucidation workflows. However, they often require an initial suggestion of the target molecule and may not be practical for de novo structure elucidation. |
67339b2df9980725cff94c52 | 2 | Other methodologies have been proposed like those by Pesek et al. that integrate data from IR, H and 13 C NMR, and mass spectra to build molecular structures, simulating the process a spectroscopist would follow. More recently, some models have been developed to process IR or H and C NMR spectra, transforming spectral data into tokenized text formats to predict molecular structures as SMILES. Other frameworks assess structural connectivity by processing 1 H and C NMR spectra, predicting substructures and assembling candidate isomers with probabilistic rankings. Additionally, DeepSAT, a CNN-based system, uses HSQC spectrum data for scaffold prediction. The NMR-TS method combines machine learning and density functional theory to automate molecule identification from NMR spectra. However, this neural network generates candidate structures without directly considering the spectra, relying on chance to predict the correct molecules. In IR spectroscopy, advancements in deep and convolutional neural networks now enable functional group identification from FTIR spectra without relying on databases or rule-based methods. However, these current approaches still face limitations, such as 1) applicability domain & reliance on extensive databases 2) need for suggestion of target molecule 3) limited consideration of multiple data modalities Our approach addresses these limitations by proposing an automated pipeline from spectra to molecular structure. This pipeline leverages the Transformer neural network architecture that can simultaneously process multiple spectroscopic data types ( 1 H-NMR, C-NMR, HSQC, COSY, IR, and MS). The Transformer's attention mechanism allows it to focus on the most relevant spectral features across different data types, enabling it to learn complex relationships between spectral inputs and molecular structures. |
67339b2df9980725cff94c52 | 3 | Furthermore, we introduce an innovative improvement cycle that allows the model to adapt to unseen chemical spaces. This iterative process enhances the model's ability to predict structures in novel domains, effectively expanding its applicability. Importantly, this improvement cycle enables our model to solve real experimental spectra despite being trained initially on simulated spectra. This capability demonstrates the model's robustness and potential for practical applications in real-world structure elucidation tasks. |
67339b2df9980725cff94c52 | 4 | This study utilizes simulated spectroscopic data across multiple modalities: 1 H NMR, C NMR, COSY, HSQC, IR spectra and mass spectrometry (MS) information. The 1 H and C chemical shifts are generated using the Scalable Graph Neural Network (SGNN). These shifts serve as the basis for both 1D NMR information and the generation of 2D spectra. Two-dimensional NMR spectra (COSY and HSQC) were reconstructed from the SGNN-predicted 1 H and C shifts. For HSQC, we employed a reconstruction logic validated against state-of-the-art simulation software. A similar rule-based approach was implemented for COSY spectra, accounting for molecular structure and H-H coupling patterns. To simulate 1 H peak splitting patterns, we developed a rule-based algorithm considering Jcouplings, producing spectra that mimic real-world 1 H NMR data. C shifts were presented without intensity values, consistent with experimental practices. IR spectra were simulated using ChemProp-IR, a directed message passing neural network. MS information, specifically molecular weights and formulas was derived from the SMILES structures using RDKit, simulating the output of highresolution LCMS measurements. While exact LCMS data was not directly processed in our workflow, our approach assumes accurate molecular weight and formula determination from MS analysis. For data handling, spectral data for each modality (except IR) is stored in CSV files containing SMILES strings, unique sample identifiers, and corresponding spectral information. IR spectral data is stored individually, with file names serving as molecular identifiers. In preprocessing, chemical shift data for 1 H and C NMR spectra are normalized by factors of 10 and 200, respectively, with this normalization also applied to the relevant dimensions in the HSQC and COSY spectra. IR spectra are down sampled to 1,000 data points along the frequency dimension. Prior to training, all data is consolidated into a single .pkl file, with spectra, SMILES strings, and sample identifiers stored in a dictionary format, ensuring efficient data retrieval during the training process. Detailed information on the reconstruction logic, simulation of coupling constants and splitting patterns, and limitations of the spectral simulations can be found in Supplementary Information Section 1. |
67339b2df9980725cff94c52 | 5 | The initial training dataset was created by randomly selecting approximately 5 million molecules from the ZINC270 database (accessed via DeepChem Python package, interface-documentation/downloads), within the 250-350 Dalton molecular weight range. This dataset was split into training and testing subsets at a 9:1 ratio. To evaluate model generalizability, a secondary dataset of 1.5 million molecules (0-500 Daltons) was extracted from PubChem. Molecules containing elements Se, Sn, As, Ge, Te, Al, Hg, Ga, Sb, Pb, Tl, Bi, Ti, Li, Zn, Na, and Pd were excluded to maintain consistency with the ZINC dataset composition. To ensure dataset uniqueness and prevent overlap, we compared the canonicalized SMILES representations of molecules from both ZINC and PubChem datasets, removing any duplicates. |
67339b2df9980725cff94c52 | 6 | Embedding Layers: Raw peaks from each modality ( 1 H, C, HSQC, COSY, and IR) are transformed into a 128-dimensional latent space. NMR spectra are normalized ( 1 H dimensions divided by 10, C by 200) and zero-padded to 64 inputs. H NMR data is provided as 2D input (chemical shift and intensity), while HSQC and COSY data are provided as 2D inputs (x and y chemical shift coordinates without intensity). |
67339b2df9980725cff94c52 | 7 | C NMR data is provided as 1D input, using only chemical shifts without intensity information. IR spectra (400-4000 cm -1 ) are discretized and down sampled to a total of 1000 datapoints. These datapoints are then embedded to match the 128-dimension input expected by the model. After embedding of each spectrum, a rectified linear unit (ReLU) activation is applied, and a mask is used to identify actual data points. Finally molecular weight and formula embeddings are concatenated to each spectrum embedding to be considered in each encoder type. Encoder: Individual spectrum encoders, each comprising 6 encoder blocks with 16 attention heads and a 4x forward expansion, process the embeddings. The resulting feature vectors are then concatenated and fed into a cross-modality encoder with an identical structure. This cross-modality encoder employs self-attention mechanisms to mix and correlate information from different spectral inputs, allowing the model to identify inter-spectral relationships. Decoder: A 6-block and 16-attention head decoder processes the encoder output along with target SMILES strings. HSQC & COSY Matching: The HungDist-NN algorithm is used for HSQC & COSY peak matching. It compares the HSQC & COSY spectra of generated analogs against the target molecules' spectra. In this process, each generated molecule's spectrum is compared to the input spectrum, and molecules are ranked according to their individual HSQC and COSY errors. These rankings are then combined to produce a final overall ranking. The analog with the lowest spectral discrepancy, as determined by the best combined ranking, is identified as the most likely correct structure. |
67339b2df9980725cff94c52 | 8 | To optimize the performance and efficiency of our MMT model, we conducted a systematic study exploring various model configurations and training dataset sizes. This analysis was crucial for determining the most effective architecture and data volume for our specific task of molecular structure prediction from spectroscopic data. We conducted a systematic study exploring three model configurations: |
67339b2df9980725cff94c52 | 9 | Large: 6 encoder and decoder layers, 16 attention heads Each configuration was trained on spectral data from 1 million molecules from the ZINC dataset for 20 epochs, using Cross Entropy loss for SMILES prediction which allowed a direct comparison of architectural complexity under controlled conditions. We also investigated the impact of training dataset size on model performance. Datasets of 100k, 1M, 2M, and 4M molecules were prepared, with training epochs adjusted to maintain a constant exposure of 20 million molecules: |
67339b2df9980725cff94c52 | 10 | Integration of Molecular Weight Loss: This stage refined the model's performance by incorporating molecular weight comparisons, leveraging the MS information. The model generates SMILES strings, from which molecular weights are calculated and compared to the target molecular weights using PyTorch's MSELoss function. This loss was normalized via min-max scaling to align with SMILES loss metrics. The molecular weight loss contribution was gradually integrated into the total loss, increasing by 1% every 50,000 steps. This gradual integration allowed the model to smoothly adapt to the new loss component without disrupting the learning process for SMILES prediction. |
67339b2df9980725cff94c52 | 11 | Training with Spectral Data Dropout: The model was trained under conditions of random spectral data omission to enhance resilience and performance in data-constrained environments. Each spectrum had a 50% chance of being omitted and replaced with zeropadded data to maintain input consistency. This stage continued to utilize SMILES and molecular weight losses. Additionally, to quantify each modality's contribution to the predictive accuracy, we performed an ablation study. The fully trained MMT model underwent single epoch fine-tuning iterations, each omitting one spectral modality. Performance comparisons between these ablated models and the complete MMT model elucidated the relative importance of individual spectral data types in molecular structure prediction. |
67339b2df9980725cff94c52 | 12 | The dataset underwent a 9:1 train-test split, with the training data further divided 9:1 to create a validation set. Training was executed in multiple five-epoch stages using four Nvidia V100 GPUs, with a batch size of 256 molecules (64 per GPU), spanning approximately seven days per stage. The PyTorch Lightning framework was utilized for efficient multi-GPU training management. The AdamW optimizer was employed with an initial learning rate of 1e -4 . Learning rate adjustment was managed by the ReduceLROnPlateau scheduler, which reduced the rate by a factor of 0.5 following two consecutive epochs without loss improvement. For the improvement cycle fine-tuning, the learning rate was adjusted to 2e -4 when using the simulated PubChem dataset. When working with the ACD Labs generated data and the experimental dataset, the learning rate was further increased to 3e -4 , a 50% spectral dropout was applied, and the molecular weight loss contribution was set to 100%. These modifications in learning rate, dropout, and loss contribution facilitated more effective fine-tuning on specific datasets. |
67339b2df9980725cff94c52 | 13 | The multinomial sampling process incorporates an adaptive mechanism to generate a specified number of unique molecules. It employs an iterative approach that dynamically adjusts the temperature parameter. Starting from an initial value of 1, the temperature is incrementally increased by 0.1 for each iteration if the desired number of unique molecules is not achieved. This process continues until either the target number of unique molecules is generated, the temperature reaches a maximum value of 3, or the iteration count hits a limit of 500. This adaptive strategy ensures a balance between diversity and computational efficiency, while guaranteeing the termination of the sampling process. The improvement cycle is a crucial component of our approach, designed to enhance the MMT model's adaptability to new chemical spaces. By iteratively generating analogs, simulating their spectra, and fine-tuning the model, we aim to improve its performance on previously unseen molecular structures. This process is particularly important for bridging the gap between simulated training data and realworld applications, where novel molecular structures are frequently encountered. |
67339b2df9980725cff94c52 | 14 | Analog Generation: Using the Mol2Mol model, we generated 10, 30, 50, or 100 analogs for each target molecule, exploring related chemical spaces without producing exact duplicates. Spectral Simulation: NMR and IR spectra were simulated for all generated analogs. Fine-tuning: The MMT model underwent fine-tuning on these datasets for 50 epochs to enhance performance on the targeted chemical space. |
67339b2df9980725cff94c52 | 15 | The improvement cycle allows the model to adapt to new chemical spaces by incorporating structurally similar molecules into the training process. By fine-tuning the model on these augmented datasets, we enhance its ability to predict structures for novel compounds not represented in the original training data. The performance quantification utilized averaged correct sample probability and averaged Tanimoto similarity (via greedy sampling). Post-fine-tuning, multinomial sampling generated three candidates per target molecule, from which the highest Tanimoto similarity was used for assessment. The PubChem evaluation expanded testing to three molecular weight ranges: 0-250, 250-350, and 350-500 Daltons, with 100 molecules in each category. This allowed assessment of model performance on molecular weights varying from the initial training data. The same improvement cycle was applied to the PubChem dataset. Additionally, we fine-tuned the ZINC-trained MMT model on the entire PubChem training set for 18 epochs, comparing its performance to that achieved through targeted improvement cycles on smaller subsets. A second improvement iteration was conducted using the checkpoint from the smallest dataset (10 analogs) testing for the iterative improvement capabilities of the model. The fine-tuning learning rate was increased to 2e -4 for this iteration, applied across all molecular weight settings in the PubChem dataset. |
67339b2df9980725cff94c52 | 16 | The Mol2Mol is a sequence to sequence transformer that produces close analogues to input molecules. This approach is essential for expanding the training data into unexplored chemical spaces, enhancing the model's ability to learn and predict in these regions. The process begins with SMILES standardization using RDKit, converting molecules to a canonical form and extracting molecular scaffolds. New molecules are then generated by modifying side chains or functional groups while retaining the core scaffold structure. To ensure practical relevance and drug-likeness of the synthesized molecules, several constraints are imposed: generation within a range of ± 100 Daltons from the template's molecular weight, a modified Lipinski's Rule of Five allowing molecules up to 550 Daltons, and a configurable Tanimoto similarity filter (default 0.3) to maintain a desired level of similarity to the target molecule. The system incorporates a scaffold hopping mechanism that shifts to a new molecular scaffold if the current one fails to produce viable candidates after a set number of iterations or if too many molecules with the same scaffold have been generated. |
67339b2df9980725cff94c52 | 17 | We designed an experiment to evaluate our models' performance across various data types, progressing from our own simulations to ACD Labs simulations and finally to real experimental data. We curated a set of 34 diverse in-house collected and publicly available molecules (see Supplementary Figure ) with all experimental spectral modalities ( 1 H-NMR, C-NMR, HSQC, COSY, IR, and MS) available. To maintain a focused fine-tuning process, we handled each molecule individually in the improvement cycle and combined the sampling performance of 3 individual runs. This individual approach was chosen to allow the model to adapt more precisely to the specific chemical features of each molecule, potentially improving its performance on challenging or unique structures at the expense of broader generalization. We first established a baseline performance using our pretrained MMT model on our simulated data. For each molecule, we conduct the improvement cycle, generating 50 analogues per target molecule (±100 Da range, max. 50 per scaffold), fine-tuning for 15 epochs with a learning rate of 3e -4 , employing a loss function with fixed molecular weight contribution of 100% and 50% spectral dropout. We reassessed performance on our simulated data, then tested on ACD Labs simulated spectra, which provided an intermediate challenge due to different underlying algorithms and error profiles. For IR simulations, we used ChemProp-IR in both our simulation pipeline and for the ACD Labs data. Finally, we evaluated the models on manually curated experimental data. Throughout all phases, we used multinomial sampling (rate 3x20) and our HSQC/COSY matching logic for molecule identification. This experiment allows us to assess the model's performance across different data sources and levels of complexity. By progressively challenging the model with data that increasingly deviates from the training simulations, we evaluate its robustness and adaptability to real-world applications. |
67339b2df9980725cff94c52 | 18 | To evaluate the robustness of our improvement cycle approach, we conducted an additional experiment simulating potential errors in initial structure assumptions. We tested the improvement cycle using slightly modified versions of the actual target molecules as starting points (shown in Supplementary Figure ). This scenario mimics real-world situations where chemists might begin with an incorrect assumption about the target molecule's structure, which is common in structure elucidation tasks. For instance, a chemist might misinterpret initial spectral data, leading to an incorrect initial structure guess. By demonstrating that our model can overcome these initial inaccuracies, we show its potential to assist in real-world structure elucidation tasks where the exact structure is unknown and initial hypotheses may be flawed. We applied this modified approach to simulated, ACD, and experimental spectra, assessing the model's ability to overcome initial structural inaccuracies and still often predict the correct molecular structure. |
67339b2df9980725cff94c52 | 19 | Our initial experiments focused on optimizing the MMT model, which processes various spectral data types including NMR ( 1 H, C, HSQC, and COSY), IR, and MS. We evaluated different model configurations and training strategies using metrics such as SMILES prediction accuracy, structural similarity, and SMILES validity of generated molecules. Larger models and datasets consistently improved performance across all metrics, leading us to select the largest model configuration and a 4 million molecule dataset for further analysis. We implemented a three-stage training strategy, progressively incorporating SMILES prediction, molecular weight loss, and spectral data dropout, which enhanced the model's overall performance. For more detailed information on the optimization process and experimental results, please refer to Supplementary Information Section 2. we evaluated the model's molecule identification accuracy using HSQC spectral matching, achieving an 89.9% accuracy with multinomial sampling, significantly outperforming greedy sampling which achieved 50.0% accuracy. These optimizations establish a robust foundation for the MMT model's application in molecular structure elucidation from spectral data. More detailed information on this experiment can be found in Supplementary Information Section 3. |
67339b2df9980725cff94c52 | 20 | To understand the relative importance of different spectral data types on model performance, we conducted an ablation study by omitting each spectral modality in turn during single-epoch fine-tuning iterations of the fully trained MMT model. We evaluated the impact on three key metrics: averaged correct SMILES sample probability, average greedy sampled Tanimoto similarity, and number of invalid molecules generated. The results, illustrated in Figure , reveal that 2D NMR data (HSQC and COSY) contribute most significantly to the model's performance. Omitting HSQC data led to the most substantial drops in correct SMILES probability (0.51 to 0.04) and Tanimoto similarity (0.82 to 0.43), while also resulting in the highest number of invalid molecules (44,847). COSY omission showed the second-largest impact, particularly evident in the notable increase of invalid molecules (38,798). It's important to note that the model's interpretation of spectral importance may differ from that of human spectroscopists. Each spectral embedding in our model includes molecular weight and formula information, enhancing its information content beyond what's visually apparent in the spectrum alone. This additional context influences the model's prioritization of different modalities. Interestingly, while 13 C NMR showed less impact on model performance, its time-consuming acquisition process in practice might make it a candidate for deprioritization in time-sensitive scenarios. Conversely, IR spectroscopy, despite showing minimal effect on performance in this study, offers rapid data collection, potentially making it valuable in practical applications where speed is crucial. Interestingly, while 2D NMR techniques (HSQC and COSY) showed the most significant impact on model performance, the time-consuming acquisition process of C NMR in practice might make it a candidate for deprioritization in time-sensitive scenarios, given its relatively lower impact on model performance. Conversely, IR spectroscopy, despite showing minimal effect on performance in this study, offers rapid data collection, potentially making it valuable in practical applications where speed is crucial. The substantial influence of 2D NMR data suggests that prioritizing HSQC and COSY spectra acquisition could significantly enhance structure elucidation accuracy, especially when balanced against time and resource constraints. |
67339b2df9980725cff94c52 | 21 | To address the challenge of the vast chemical space that no model can be trained on entirely, we integrated an improvement cycle that activates when the model encounters an unfamiliar region not covered in the training data. This cycle employs a generative model designed to suggest structurally similar molecules within the unexplored chemical space, allowing for the creation of a fine-tuned dataset tailored to these novel regions. Coupled with this process is a data generation pipeline, which includes the SGNN network for generating 1 H and C NMR spectra and rule-based algorithms for reconstructing HSQC and COSY spectra, as well as for calculating coupling constants in 1 H NMR spectra. Furthermore, IR spectra are generated using a message-passing neural network. For mass spectrometry (MS) data, we calculate the exact molecular weight from the SMILES representation of each molecule using RDKit, simulating the molecular ion peak that would be observed in highresolution MS. This comprehensive approach ensures that all relevant spectral modalities, including MS data, are represented in the fine-tuning dataset, enhancing the model's ability to adapt to new chemical spaces. |
67339b2df9980725cff94c52 | 22 | We initially tested this improvement cycle on a test set from the ZINC dataset to determine if further improvements could be achieved beyond the pretrained network. Figure presents the averaged Tanimoto similarity (a) and averaged correct sample probability (b) results for the ZINC test data before and after fine-tuning with different numbers of generated analogs. The multinomial sampling (MNS) approach, generating three candidates per target molecule, demonstrates remarkable effectiveness, identifying up to 96% of correct molecules within the top 3 candidates. For the MNS, we employed a molecular weight filter for the sampling process, accepting only molecules that fulfill these requirements. Greedy sampling also shows robust performance, correctly identifying up to 78% of molecules after fine-tuning. Model performance improves with up to 30 training analogs but plateaus or slightly declines with 50 or 100 analogs. This may result from the Mol2Mol model's tendency to generate up to 30 analogs per scaffold before switching, potentially impacting analog quality and finetuning effectiveness. For this experiment, we set the parameter for the number of samples per scaffold to 30. While the number of samples per scaffold is adjustable, we did not further investigate this parameter in the current study. |
67339b2df9980725cff94c52 | 23 | We expanded our testing to the PubChem dataset to explore the model's capabilities across varied molecular weights and chemical structures beyond the initial training set. We curated three sets of 100 molecules from PubChem, categorized into molecular weight ranges: 0-250 Da (see Figure ), 250-350 Da, and 350-500 Da (see Supplementary Figure ). |
67339b2df9980725cff94c52 | 24 | We applied the improvement cycle methodology, previously used for the ZINC dataset, to PubChem. This involved generating molecular analogs via the Mol2Mol model, simulating spectral data, and fine-tuning the model (details in Methods section). For comparison, we also fine-tuned the ZINC-trained MMT model on the entire PubChem training set (referred to as PC-FT), allowing us to assess the effectiveness of our targeted improvement cycles against a more comprehensive fine-tuning approach. |
67339b2df9980725cff94c52 | 25 | Iterative application of the cycle on the 10-analog generation model showed further improvements in model accuracy (Supplementary Figure ). Comprehensive visualizations of the model's improvement and chemical space exploration are provided in Supplementary Information Section 5, including t-SNE plots and examples of molecules from different weight ranges and their generated analogs. |
67339b2df9980725cff94c52 | 26 | Our evaluation across molecular weight ranges demonstrates the model's adaptability and the improvement cycle's effectiveness in enhancing performance on diverse structures. The cycle significantly improves identification accuracy, with notable results even when using just 10 analogs for weight ranges covered in initial training. For the ZINC dataset (250-350 Da), the IC increased perfect Tanimoto matches from 50% to 73% with 10 analogs, and up to 78% with 30 analogs. In the PubChem dataset (0-250 Da), the IC improved perfect matches from 14% to 45% with 10 analogs, surpassing the 44% achieved by the model fine-tuned on the entire PubChem dataset. Multinomial sampling consistently outperforms greedy sampling in identifying correct molecules, achieving up to 96% and 68% accuracy within the top 3 candidates for the ZINC and PubChem dataset, respectively. However, selecting the single most accurate candidate remains a challenge. The following section explores an additional step to address this. |
67339b2df9980725cff94c52 | 27 | This experiment evaluates the performance of the MMT model using a selected set of 34 molecules for which all experimental data modalities ( 1 H, C, HSQC, COSY, IR, and MS) were available. The peaks of all real experimental data used in this study was manually peak picked to ensure accuracy and consistency. A preliminary study (Supplementary Section 2) revealed that multinomial sampling (MNS) combined with peak matching significantly outperformed greedy sampling, increasing identification accuracy from 50% to 90% (see Supplementary Figure ). Despite challenges in differentiating similar structures (examples in Supplementary Figure ), we adopted MNS with spectral error ranking, replacing greedy sampling for this experiment. |
67339b2df9980725cff94c52 | 28 | For each target, we sampled molecules using multinomial sampling (MNS with sample size 3x 20) and applied molecular weight filters to ensure the generated molecules matched these key properties of the target compound. After sampling we employed the HSQC matching methodology developed in our previous research, using the HungDist-NN matching algorithm to score and rank the set of sampled molecules. Additionally, we applied the same point matching methodology for the COSY spectrum and investigated its impact on molecule ranking. Evaluation uses three ranking methods: COSY, HSQC, and combined HSQC & COSY, with top-k accuracy calculated for k = 1, 3, 5, 10, 20 and "Total" representing all generated molecules out of the maximum of 60 sampling options. Results for the MMT model, illustrated in Figure , show near-perfect accuracy on our simulated data (a-c), with already top-3 accuracy reaching 100% across all ranking methods. Performance on ACD Labs simulated data (d-f) remains strong, with top-3 accuracy of 67% and reaching 94% for top-20. Experimental data (g-i) shows similar performance, with top-3 accuracy between 62% and a total top performance of 94% over all sampled molecules. Notably, the combined HSQC & COSY ranking outperforms individual rankings across all data types, suggesting enhanced prediction accuracy. For instance, in experimental data, the combined ranking achieves 91% top-10 accuracy compared to 88% for COSY or HSQC alone. |
67339b2df9980725cff94c52 | 29 | The baseline performance of the pretrained MMT model without the improvement cycle, detailed in Supplementary Information Section 6, demonstrates the model's initial limitations: solving only 58% of our simulated data, 16% of ACD Labs simulated data, and a mere 3% of real experimental data. In contrast, the improved model achieves high accuracy on experimental data, highlighting the effectiveness of our improvement cycle. This dramatic enhancement, despite training solely on simulated spectra, underscores that comprehensive chemical space coverage is more critical than precise training data accuracy. The model's ability to adapt to various spectral simulation methods and real experimental data showcases its robustness and generalizability, crucial for diverse research environments. These results demonstrate the model's practical applicability in real-world structure elucidation tasks, with the potential to significantly accelerate and improve the structure determination process in chemistry laboratories. Furthermore, the model's adaptability to discrepancies between simulated and experimental data suggests that additional fine-tuning with real experimental data could yield even greater improvements, opening avenues for further enhancing its performance in practical settings. |
67339b2df9980725cff94c52 | 30 | Comparing the total number of correctly identified molecules, we observed the following: For our simulated data, the model identified 34 out of 34 (100%) molecules with correct starting structures, and 29 out of 34 (85%) with modified starting structures. With ACD simulated data, 31 out of 34 (94%) molecules were found using correct starting structures, compared to 22 out of 34 (65%) with modified starting structures. For experimental data, the model identified 31 out of 34 (94%) molecules correctly with accurate initial guesses, versus 19 out of 34 (56%) with modified starting points. These results demonstrate that while there is a decrease in performance when starting with modified structures, the model still maintains a substantial ability to identify correct molecules. This robustness is valuable in real-world applications where initial structural assumptions may not always be entirely accurate, showcasing the Mol2Mol network's ability to explore relevant chemical space and allow the MMT model to overcome initial inaccuracies in many cases. A figure showing all the wrong starting guesses of the molecules is presented in Supplementary Figure . |
67339b2df9980725cff94c52 | 31 | The model's ability to maintain good performance even with slightly incorrect initial structures has significant implications for real-world applications. In practice, chemists often begin structure elucidation with incomplete or partially incorrect hypotheses. Our model's robustness in these scenarios suggests it can serve as a powerful tool to refine and correct initial structural guesses, potentially reducing the iterative cycles typically required in structure determination. This capability could be particularly valuable in analyzing complex natural products or in drug discovery processes where rapid and accurate structure elucidation is crucial. This visualization allows us to identify which structural features the model is most certain about and which areas might require further refinement, providing a window into the model's intuition process. Similar analyses for our simulated data and ACD Labs predictions are shown in Supplementary Figure and Supplementary Figure . |
67339b2df9980725cff94c52 | 32 | These explainability features provide practical benefits for structure elucidation tasks. By visualizing the model's confidence in structural components, chemists can focus on uncertain aspects, guiding targeted experimental work like selective 2D NMR or chemical derivatization. For multiple suggested structures, confidence visualization aids in prioritizing hypotheses, streamlining the elucidation process. The correlation between prediction probabilities and spectral reconstruction errors offers a metric for assessing prediction reliability, helping chemists decide when to trust the model's suggestions or seek additional experimental evidence. To facilitate practical application of these insights, we have implemented the full improvement cycle workflow as a GUI in the form of an HTML website. This interface includes probability plotting for suggested molecules and allows users to compare simulated spectra of generated molecules with data. The code and user manual are provided in the Supplementary Information Section 7 with explanatory screenshots in Supplementary Figure -Supplementary Figure , offering a powerful tool to elucidate structures based on our MMT model. |
67339b2df9980725cff94c52 | 33 | To leverage these insights in practical applications, we have implemented the full improvement cycle workflow together with the probability plottings of the suggested molecules. This implementation also includes an option for users to compare the simulated spectra of the generated molecules with the experimental spectra. This feature should aid in investigating potential wrong assignments and mistakes of the model, providing a powerful tool for chemists to critically evaluate and refine the model's predictions. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.