Filename
stringlengths 22
64
| Paragraph
stringlengths 8
5.57k
|
|---|---|
Processed_From_gas_to_galaxies.txt | Figure 7. The H I mass function for different mor- phological types. Shown are Schechter fits (solid lines) and the overall H I mass function for the to- tal sample (dashed line). From Zwaan et al.[106]. |
Processed_From_gas_to_galaxies.txt | Virgo cluster the H I deficiencies and H I to opti- cal sizes are notably different (respectively larger and smaller) than in the less dense environment of the Ursa Major cluster and the field. |
Processed_From_gas_to_galaxies.txt | The velocity structure and spatial distribution of H I detected galaxies near clusters suggests that infall is still going on and that clusters are still capturing galaxies which fall in along large scale structure filaments. Evidence such as the H I structures around S0 galaxies and the stripped gas in NGC 4522 (Fig. 10 and Fig. 11) are cru- cial in pinning down these processes and SKA will be able to see such structure over a considerable range of redshift (i.e. evolution). |
Processed_From_gas_to_galaxies.txt | Figure 11. NGC 4522 in the Virgo cluster from Kenney et al. [54]. The H I contours are displaced from the optical disk and clearly show the sweep- ing effect of ram-pressure stripping. |
Processed_From_gas_to_galaxies.txt | taneous bandwidth of 500 MHz) will be able to see NGC 3359–like structures to z = 0.15 opening the exciting prospect of being able to probe the environment of several million galaxies and pin down the merger/accretion characteristics at the present time over a range of galaxy densities. The resolution and sensitivity of such a survey will be comparable to that of the images shown in Fig. 9 and Fig. 10. The volume covered by such a survey is ∼ 15,000 Mpc3 and will detect ∼ half a million galaxies, providing a detailed H I image, not only of the galaxies but also of their H I environments. Only with such data can the evolution of galax- ies, in particular the mechanisms for converting gas into stars in and around galaxies be studied. |
Processed_From_gas_to_galaxies.txt | Figure 12. Top: H I deficient fraction in bins of projected radius from the cluster center for the superposition of all the H I deficient clusters. Bot- tom: H I deficiency versus projected radius from the cluster center. Small dots show the radial variation of H I deficiency for individual galaxies, while the arrows indentify non detections plotted at their estimated lower limit. Large dots are the medians of the binned number distribution. From Solanes et al. [80]. |
Processed_From_gas_to_galaxies.txt | Figure 13. A comparison of H I properties of galaxies in the Ursa Major and Virgo clusters as a function of projected distance to the cluster cen- ter. Upper panel: ratio of H I to optical diameters. Lower panel: overall H I deficiency as a function projected distance. |
Processed_Systematic_uncertainties_in_NLOPS_matching.txt | The MC@NLO and MEPS@NLO methods, as implemenated in the Monte-Carlo event generator framework SHERPA, are used to estimate the perturbative and non-perturbative uncertainties in various processes such as dijet production and the production of a W boson in association with (multiple) jets. |
Processed_Systematic_uncertainties_in_NLOPS_matching.txt | c(cid:13) Copyright owned by the author(s) under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike Licence. |
Processed_Systematic_uncertainties_in_NLOPS_matching.txt | Being largely stimulated by the need for higher precission of theoretical predictions in both Standard Model analyses and new physics searches at the LHC, the simulation of higher-order QCD corrections in Monte Carlo event generators has seen vast improvements in recent years. To this end, two lines of development have been followed. In the MEPS approach [1, 2, 3, 4, 5, 6] higher-order tree-level matrix elements of successive final state parton multiplicity are merged into an inclusive sample, offering both leading-order accuracy for the production of hard partons and re- taining the overall resummation of scale hierarchies through the parton shower at the same time. On the other hand NLOPS approaches, introduced as either MC@NLO [7] or POWHEG [8, 9], work on a single parton multiplicity elevating its accuracy to next-to-leading order. Both methods have been shown to be automatable [10, 11] within the SHERPA event generator framework [12]. Thereafter, it was sought to recombine both lines of development. In a first step, called the MENLOPS prescrip- tion, the NLOPS and MEPS methods have been combined using the NLOPS’ NLO accuracy for the inclusive process supplementing it with higher-order tree-level matrix elements in an MEPS fash- ion [13]. In second step multiple NLOPS processes of successive parton multiplicity are combined, elevating the accuracy of the MEPS method to next-to-leading order, dubbed MEPS@NLO [14, 15]. In the following both the NLOPS and MEPS@NLO methods are summarised. Particular emphasis is put on both methods’ major accompishments with respect to standard leading order computations: its increased theoretical accuracy expressed through reduced perturbative uncertainties. |
Processed_Systematic_uncertainties_in_NLOPS_matching.txt | Figure 1: Rapidity (left) and azimuthal (right) separation of the two leading jet pp ATLAS data [16]. |
Processed_Systematic_uncertainties_in_NLOPS_matching.txt | Figure 2: Left: Inclusive jet cross section in pp 2-jet ratio in dependence on the scalar transverse momentum sum of all jets in pp to CMS [18]. |
Processed_Systematic_uncertainties_in_NLOPS_matching.txt | of perturbative (µR, µF , µQ) as well as non-perturbative uncertainties due to the multiple interac- tion model in an MC@NLO implementation of inclusive and dijet production [20]. In all cases, the perturbative uncertainties for observables described at NLO accuracy are greatly reduced while the parton shower resummation provides the correct description when large hierarchies of scales in t are present. At the same time, there are observables/regions where the uncertainty on the modelling of the soft structure of the event is non-negligible. |
Processed_Systematic_uncertainties_in_NLOPS_matching.txt | Therein an MC@NLO description of an n parton multiplicity is restricted to have its emission produced at a jet measure Q smaller than Qcut. The region with Q > Qcut is then filled with an MC@NLO for the n + 1 parton process. To restore the correct resummation with respect to the n parton process to at least parton shower accuracy its Sudakov form factor ∆(PS) is inserted. The overlap with similar terms in ¯B(A) n+1 is removed with the term in the braces on third line. A multijet merged description is then achieved by iteration eq. 3.1. |
Processed_Systematic_uncertainties_in_NLOPS_matching.txt | Figure 3: Cross section as a function of the inclusive jet multiplicity in pp ATLAS data [16]. |
Processed_Systematic_uncertainties_in_NLOPS_matching.txt | boson in pp collisions at √s = 7 TeV with the ATLAS detector, Phys.Rev. D85 (2012), 092002, [arXiv:1201.1276 [hep-ex]]. |
Processed_Systematic_uncertainties_in_NLOPS_matching.txt | Cross Sections in pp Collisions at √s = 7 TeV, Phys.Lett. B702 (2011), 336–354, [arXiv:1106.0647 [hep-ex]]. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | We present new fluorine abundance estimations in two carbon enhanced metal- poor (CEMP) stars, HE 1429–0551 and HE 1305+0007. HE 1429–0551 is also enriched in slow neutron-capture process (s-process) elements, a CEMP-s, and HE 1305+0007 is enhanced in both, slow and rapid neutron-capture process elements, a CEMP-s/r. The F abundances estimates are derived from the vibration-rotation transition of the HF molecule at 23358.6 ˚A using high-resolution infrared spectra obtained with the Immersion Grating Infrared Spectrometer (IGRINS) at the 4m-class Lowell Discovery Telescope. Our results include a F abundance measurement in HE 1429–0551 of A(F) = +3.93 ([F/Fe] = +1.90) at [Fe/H] = −2.53, and a F upper limit in HE 1305+0007 of A(F) < +3.28 ([F/Fe] < +1.00) at [Fe/H] = −2.28. Our new derived F abundance in HE 1429–0551 makes this object the most metal-poor star where F has been detected. We carefully compare these results with literature values and state-of-the- art CEMP-s model predictions including detailed AGB nucleosynthesis and binary evolution. The modelled fluorine abundance for HE 1429–0551 is within reasonable agreement with our observed abundance, although is slightly higher than our observed value. For HE 1429-0551, our findings support the scenario via mass transfer by a primary companion during its thermally-pulsing phase. Our estimated upper limit in HE 1305+0007, along with data from the literature, shows large discrepancies compared with AGB models. The discrepancy is principally due to the simultaneous s- and r-process element enhancements which the model struggles to reproduce. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | that a considerable fraction of VMP stars exhibit high carbon enrichment. These Carbon Enhanced Metal-Poor (CEMP) stars are defined as those with [C/Fe] 1 > 0.7 (Aoki et al. 2007) and the number of CEMP stars increases with decreasing metallicities. For example, at [Fe/H] = −2.0, the CEMP fraction reaches ∼20%, while at [Fe/H] = −4.0, this fraction rises up to ∼ 80% (Placco et al. 2014). In addition to carbon, other element enhancements are found in the stellar atmospheres of these peculiar objects such as light elements and neutron-capture elements. The abundance distribution of these elements in CEMP stars normally follow a few distinctive patterns, which have a direct connection to the enrichment mechanisms that operated in the early Galaxy. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | CEMP-s stars are CEMP stars that satisfy the criteria [Ba/Fe] > +0.5 and [Ba/Eu] > 0; • CEMP-r stars show abundance enrichments of [Eu/Fe] > +1.0 and [Ba/Eu] < 0; • CEMP-s/r are CEMP-s stars, [Ba/Fe] > +0.5 and [Ba/Eu] > 0, also enriched with r-process elements, i.e. [Eu/Fe] > +1.0; • CEMP-no stars do not exhibit enhanced abundances in r− and s-process elements, i.e. [Ba/Fe] < +0.5. Nomoto et al. (2013) summarised the different channels for the formation of CEMP stars. Some of these channels require a first or second generation supernova to pollute the gas cloud that forms the CEMP star with carbon and other metals. Here though, we focus on the mass transfer scenario which takes place in a binary system. According to this scenario the CEMP star receives its carbon and heavy elements from a previous Asymptotic Giant Branch (AGB) companion. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | parameters and final conditions of the binary system, as well as abundance predictions for several elements in those CEMP-s stars, including fluorine. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Fluorine is a particularly rare and fragile element. With just a single stable isotope, 19F, fluorine is the least abundant of the elements from C to Ca. The synthesis of 19F requires a series of specific conditions, and once formed, it can be easily destroyed in the stellar interiors (Woosley & Haxton 1988; Jorissen et al. 1992; Forestini et al. 1992; Mowlavi et al. 1996; Meynet & Arnould 2000; Lugaro et al. 2004; Palacios et al. 2005; Cristallo et al. 2009). This makes fluorine difficult to produce and extremely sensitive to the environment of its formation site. Three formation sites have been proposed as principal sources of 19F. AGB stars produce fluorine through complex nuclear reactions starting from 14N and captures of protons, neutrons and α particles in the reaction 14N(α, γ)18F(β+)18O(p, α)15N(α, γ)19F (Fores- tini et al. 1992). The F content is then released from the star by stellar winds (Forestini et al. 1992; Mowlavi et al. 1996, 1998; Lugaro et al. 2008). AGB stars are the only ob- servationally confirmed site of F production (Jorissen et al. 1992; Werner et al. 2005; Zhang & Liu 2005; Pandey 2006). Other formation sites are theoretically predicted to make fluorine, however, these have not yet been confirmed by observations. Wolf-Rayet stars can also produce fluorine in a similar way to AGB stars. Although production strongly depends on the metallicity and rotation of the star (Meynet & Arnould 2000; Palacios et al. 2005; Prantzos et al. 2018). Type II supernovae can synthesize F via neutrino spallation or neutrino process, ν-process (Kobayashi et al. 2011b). High-energy neutrinos ντ ejected during the core-collapse interact with the 20Ne outer layer producing 19F (Woosley & Haxton 1988; Woosley & Weaver 1995). |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | The chemical evolution of fluorine is highly uncertain. Most of the tests for F production are provided by observed F abundances in stars at or near solar metallicities (see e.g., Jorissen et al. 1992; Cunha et al. 2008; Abia et al. 2010; Recio-Blanco et al. 2012; J¨onsson et al. 2014b; Guer¸co et al. 2019; Ryde et al. 2020). This metallicity bias is due to the difficulties to observe the HF molecular absorption line, the most used absorption feature for the diagnostic of F abundances, in VMP stars. At very low-metallicities, F has been poorly studied with just two reported measurements below [Fe/H] < −2: HE 1305+0132 by Schuler et al. (2007)3 and HD5223 by Lucatello et al. (2011) (some other mea- surements and upper limits are also available in literature, including carbon-normal halo field stars at slightly higher metallicities - e.g., Lucatello et al. 2011; Li et al. 2013). How fluorine is made in the early universe remains unclear. Therefore, studying F in CEMP stars is a direct probe to this question. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Figure 1. Observed spectra (black circles) and best fitting synthetic spectra (green line) for HE 1428-0551 (upper) and HE 1305+0007 (lower) near the HF line. Note that while the HF line is weak, most of the absorption lines are well fitted. In the lower panel, the HF line region is higher than the continuum suggesting issues with the telluric division (see Section 4). |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | the target selection and observations. Section 3 includes the data reduction and analysis. We present our results and dis- cussion in Section 4. We finalize in Section 5 summarizing our conclusions. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Table 1. Star ID and stellar parameters from literature data compilation by Abate et al. (2015b) in columns 1-5; our observed fluorine abundance in columns 6-7; modelled abundances from Abate et al. (2015b) in columns 8-9. Additionally, we have included stellar parameters, observed A(F) and [F/Fe] from Lucatello et al. (2011), and the predicted A(F) from Abate et al. (2015b) for in common stars. Note that F abundance in HE 1305+0007 has been also analysed by Lucatello et al. (2011) (see text). |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | CS 22942–019 - CS 29497–034 - CS 30314–067 - CS 22948–027 - CS 29502–092 - CS 30322–023 - HE 1152–0355 - HE 1305+0007 - HD 187861 - HD 5223 - HD 122563 - ♣ The adopted solar abundance for F is from Grevesse et al. (2007), A(F)(cid:12) = 4.56. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | periods. The comparison of this model grid to a sample of 67 CEMP-s stars was used to determine the best-fitting models for the observed chemical abundances, surface gravity, and period (when available) of each individual star of that sample (see Abate et al. 2015a,b, for further details). We adopted those stellar parameters (Teff, logg, [Fe/H]) and the F abundance predictions of each star from Abate et al. (2015b, and references therein), to generate synthetic spectra with the local thermodynamic equilibrium (LTE) stellar line analysis program MOOG (Sneden 1973, version 2014). Using the vibration-rotation transition of the HF molecule line near 2.3 µm, we identified potential targets as those showing a synthetic HF absorption line stronger than ∼5% relative to the continuum. From these potential targets, we selected two stars to perform our follow-up observations, HE 1429–0551 and HE 1305+0007. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | HE 1429–0551 is a CEMP-s star with no reported abundances of Eu. We adopted the stellar parameters compiled from the literature by Abate et al. (2015b): Teff = 4700, log g = 1.50, [Fe/H] = −2.53 from Aoki et al. (2007). A barium abundance of [Ba/Fe] = +1.57 is also reported by Aoki et al. (2007). Models by Abate et al. (2015b) predicts a F content in HE 1429-0551 of A(F)Abate = +4.40. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Both objects appear to be binaries based on previ- ous radial velocity variations. However, orbital solutions have not yet been published. HE 1429–0551 exhibits ra- dial velocity variations above the 3σ level, but no clear trend is observed to support orbital motion (Jorissen et al. 2016). HE 1305+0007 radial velocities were first reported by Goswami et al. (2006), and later by the Gaia Collaboration et al. (2018), both studies have uncertainties > 1km s−1 and a velocity discrepancy > 10km s−1 . HE 1305+0007 has been flagged as a binary candidate (Arentsen et al. 2019). |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | The CEMP stars HE 1429–0551 and HE 1305+0007 were observed using the Immersion Grating Infrared Spectrom- eter (IGRINS) (Mace et al. 2018; Park et al. 2014; Yuk et al. 2010), mounted on the Lowell Discovery Telescope at the Lowell Observatory. Observations were performed under service mode on 09 and 14 April 2019. The 0.63” slit provided spectral resolution of R ∼ 45,000, and covered the infrared H and K bands from 1.45 to 2.45 µm. Although the instrument has a large spectral coverage, for the purpose of this work we use a small fraction of the spectra at ∼2.3µm. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | HE 1429–0551 (cid:63) The total value is the quadrature sum of the individual abundance dependences. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | served twice during the first night and once during the sec- ond night with exposure times of 450 and 480 seconds per frame, respectively. The signal-to-noise ratios (S/N), exceed 150 per resolution element in the wavelength-region of in- terest, at 2.3 µm, in each star. In addition to the science observations, rapidly rotating hot stars were observed each night, close in time to our program targets, in order to ob- tain telluric spectra. The telluric spectra are divided into the science spectra of HE 1305+0007 and HE 1429–0551 in order to remove the telluric absorption lines. The program stars and log of observations are listed in Table 2. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Wavelength-calibrated spectra were produced using the IGRINS reduction pipeline package PLP2 (Lee et al. 2017). PLP2 performs the flat fielding, background removal, order extraction and wavelength calibration. IRAF4 tasks were used to combine the individual spectra into a single spectrum and to normalise the continuum. Inspection of the spectrum of HE 1305+0007 (See Figure 1) indicates that there was likely an issue with telluric removal in the vicinity of the HF line, i.e., the normalized flux exceeds 1.0. We refer to the effect of this last issue on our F abundance estimation in HE 1305+0007 in Section 4. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | The adopted stellar parameters (Teff , log g , [Fe/H] ) for our sample and its literature source are described in , Section 2.1 and Table 1. The microturbulent velocity ξt was determined using the relation ξt = 4.2 − (6 × 10−4 Teff ) (Mel´endez et al. 2008). The abundance dependence due to uncertainties in the adopted stellar parameters for HE 1429–0551 are presented in Table 3. The reason that we adopted the stellar parameters from the literature instead of derive them directly from the spectra is that 1) Adopting new stellar parameters would lead to inconsistency with the values used in the Abate predictions, and 2) new stellar parameters cannot be easily obtained from the IGRINS spectra alone. This is due to the lack of Fe lines in the wave- length coverage of the instrument used for the estimation of Teff and logg through excitation and ionization equilibrium, respectively. Alternative, photometric temperatures are feasible, but these are C-rich objects and caution should be exercised when using the standard calibrations which are generally based on C-normal objects. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Figure 2. Observed and synthetic spectra near the HF 23358.329˚A line for HE 1429–0551 (upper) and HE 1305+0007 (lower). In the upper panel, the solid green line is our best fit for HE 1429–0551 with A(F) = 3.93. The uncertainty region of our fit is shown in shaded green area and delimited by a dotted green line with A(F) = 3.93±0.4. The predicted F abundance from Abate et al. (2015b) of A(F)Abate = 4.40 is shown in the solid blue line. In the lower panel, our upper limit for HE 1305+0007 of A(F) = 3.28 is the solid green line and the predicted value from Abate et al. (2015b) of A(F)Abate = 4.52 is the solid blue line. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | and is considered a reliable indicator of F abundances (Abia et al. 2009; Lucatello et al. 2011; J¨onsson et al. 2014a). We adopted the most recent data for the HF(1ˆa ˘A¸S0) R9 molec- ular line from J¨onsson et al. (2014a, and references therein); excitation potential χ = 0.227 and log g f = -3.962. These HF features were incorporated in our line list and are com- patible with the partition function built into MOOG. By doing this we ensure that we are using energies consistent with the partition function integrated in MOOG, avoiding over-estimations in the F abundances previously reported in literature (e.g., Lucatello et al. 2011; D’Orazi et al. 2013; J¨onsson et al. 2014b). We adopted additional molecular data from Goorvitch (1994) and Sneden et al. (2014) for CO and CN, respectively. For the fitting of CO and CN molecular ab- sorption in the spectral region nearby HF, we adopted the observed C abundances from Aoki et al. (2007) for HE 1429– 0551, and Goswami et al. (2006) and Beers et al. (2007) for HE 1305+0007, then we adjust O and N in order to fit our observed spectra in these stars. Our best fitting synthetic spectra are shown in Figure 1 and 2. Note that the 12C17O line at 23358.237 ˚A, is slightly blended with the blue wing of the HF line under study. However, this partial blending does not affect significantly the fluorine measurements from the HF(1ˆa ˘A¸S0) R9 molecular line (Lucatello et al. 2011). |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Our derived fluorine abundances are: A(F) = +3.93 ([F/Fe]5 = +1.90) for HE 1429–0551 and A(F) < +3.28 ([F/Fe] < +1.00) for HE 1305+0007. The adopted metallicity on these CEMP stars are [Fe/H] = −2.53 and [Fe/H] = −2.28 for HE 1429–0551 HE 1305+0007, respectively. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | in our derived upper limit is expected to be small. The predicted strength of the HF line at very low-metallicities is weak, which is the principal reason for the high S/N ratio in our observations. We note that there is a ∼ 3% excess in the normalized flux of HE 1305+0007 (see lower panel in Figure 2). If there was an HF feature, we assume that it would have been detected if it were stronger than ∼ 5% relative to the continuum. Moreover, the typical error in the observations of F abundances is ∼0.4 dex. This error value would account for this telluric effect if the line were observed, without substantially changing our results. Naturally, more observations in HE 1305+0007 will help to clarify the F abundance in this star. For the case of HE 1429–0551, the telluric contamination was sufficiently corrected. Despite the weakness of the line (3.5% relative to the continuum), a reduced χ2 value of 2.2 in the region of interest supports our F measurement in HE 1429-0551. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | The study of F involves several sources of uncertainty. We introduce the general uncertainties in both the nu- cleosynthesis and observations as follows. The production of the 19F isotope in AGB stars takes place in the He intershell layer through complex nuclear reactions. Once formed, fluorine is dredged up to the surface during the third dredge up (TDU) along with C and s-process ele- ments (Forestini et al. 1992; Mowlavi et al. 1996, 1998; Lugaro et al. 2004). As mentioned before, the principal reaction pathway for the production of 19F in AGB stars is et al. 1992). The protons needed in the reaction are supplied via 14N(n, p)14C, which in turn is provided with neutrons by the 13C(α, n)16O reaction. This last neutron source, 13C(α, n)16O, also provides neutrons for the s-process and its efficiency depends on the amount of 13C present in the He intershell. Some 13C existent in the intershell comes from CNO cycling in previous interpulse phases, although it is not enough to achieve the AGB s-process element enhancements. As an additional source of 13C, some protons can be partially mixed into the intershell producing 13C from the abundant 12C in the region through the reaction 12C(p,γ)13N(β+ν)13C (Iben & Renzini 1982a). This reaction generates a thin layer abundant in 13C and 14N at the top of the intershell, namely the ’13C pocket’ (Iben & Renzini 1982b), which enrichment will depend on the amount of protons injected into the intershell, the partial mixing zone (PMZ). The production of F and its relation to the ’13C pocket’, suggest a positive correlation between abundances of F and s-process elements at the surface of the AGB star (see Section 4.3). However, some nuclear reaction rates are still not well known. For example, the 19F(α, p)22Ne reaction is responsible for the destruction of 19F. Although this reaction is one of the most important parameters for a proper theoretical analysis of F nucleosynthesis in AGB stars, it is also one of the principal uncertainties in the nuclear reaction rates (Lugaro et al. 2004). Other sources of uncertainty in the nucleosynthesis of 19F include the shape of H profile of the partial mixed zone, convective mixing and mass loss. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Figure 3. Our fluorine estimations along with literature values for A(F) vs. [Fe/H]. Our fluorine measurement is shown as the filled orange circle for HE 1429–0551 and our upper limit in filled orange triangle for HE 1305+0007. In blue, HE 1305+0132 from Schuler et al. (2007); in green, HE 1152–0355, HD 5223, and other upper limits from Lucatello et al. (2011); in red, HD 135148, HD 110281, and other upper limits from Li et al. (2013). The black dashed line shows the differences in the adopted [Fe/H] and the upper limit for the star in common with Lucatello et al. (2011), HE 1305+0007. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Schuler et al. (2007) measured F for the first time in a CEMP(-s) star, HE 1305+0132. This was the most metal-poor star where a F detection has been reported with A(F) = +4.96 ([F/Fe] = +2.90) at [Fe/H] = −2.5, the highest by over ∼ 1 dex from all other literature data at very low-metallicities. Lugaro et al. (2008) showed that the observed C and F abundances by Schuler et al. (2007) in HE 1305+0132 can be justified by the binary mass transfer scenario of CEMP-s stars. Their model was able to reproduce the abundances by considering: 1) accretion of 3 − 11% of the mass loss from the companion star during the AGB phase; 2) dredged-up intershell material, no less than ∼ 0.2 M(cid:12), into the convective envelope by TDU. However, the analysis by Schuler et al. (2007) used older HF molecular transition data from Jorissen et al. (1992) with an excitation potential χ = 0.48 eV. This value is not compatible with the partition function in the MOOG program (Sneden 1973), leading to an offset in the reported F abundance of ∼ +0.3 dex. (See e.g., Lucatello et al. 2011; J¨onsson et al. 2014a). Moreover, the adopted metallicity for HE 1305+0132 ([Fe/H] = −2.5 ±0.5) by Schuler et al. (2007) was derived by Beers et al. (2007) using mid-resolution optical and near-IR spectra. Later, Schuler et al. (2008) reported a revised, and higher, metallicity of [Fe/H] = −1.92 dex for HE 1305+0132 using high-resolution optical spectra, 0.6 dex higher than previously estimated. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Figure 4. A(F) versus [Fe/H] for our data (orange symbols) and Lucatello et al. (2011) (green symbols). The Abate et al. (2015b) predictions are presented in magenta x symbols. For each star, the observed and theoretical abundances are connected by a black dashed line. Note the differences in the adopted [Fe/H] between Abate et al. (2015b) and Lucatello et al. (2011) (see text for details). HE 13005+0007 is in common between our sample and Lucatello et al. (2011), and our adopted [Fe/H] for this star also differs from their work. The fluorine abundance measured in HD 5223 by Lucatello et al. (2011) are well reproduced by the theoretical model. All the upper limits are substantially overesti- mated by Abate et al. (2015b), including HE 1305+0007. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | A more metal-rich model atmosphere contains outer layers cooler than the previous metal-poor model atmosphere used in Schuler et al. (2007). This metallicity difference can induce a lower F abundance in the measurement of HE 1305+0132 by Schuler et al. (2007), comparable to those [F/Fe] abundances found in other CEMP objects and the new values obtained in this paper. Additionally, our new F measurement in HE 1429–0551 makes this object the most metal-poor star where F has been detected. (Upper limits have been obtained at similar metallicity (Lucatello et al. 2011)). |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | while Lucatello et al. (2011) adopted stellar parameters from Beers et al. (2007): Teff = 4560, log g = 1.0 and [Fe/H] = −2.5. A re-analysis of our F detection in HE 1305+0007 using the stellar parameters adopted by Lucatello et al. (2011) generates an upper limit of A(F) < 3.11, [F/Fe] < 1.05. This differs from our original analysis by ∆A(F) = +0.17 and ∆[F/Fe] = −0.05 dex. This is principally due to the high sensitivity of the adopted absorption HF line to temperature and its moderate sensitivity to metallicity. The differences in the model atmosphere used in both analyses also have an effect in the F measurements. Lucatello et al. (2011) used spherical C-enhanced models which contribute to slightly higher F abundances for HE 1305+0007 in their study (Lucatello et al. 2011, see Section 3). Additionally, issues in the removal of telluric features near the HF line described before, could also affect our derived F upper limit in HE 1305+0007. However, our upper limit is comparable with that upper limit derived by Lucatello et al. (2011) in this star. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | More recently, Li et al. (2013) derived F abundances and upper limits in 7 carbon-normal metal-poor giant stars in the halo: two measurements and five upper limits (See Figure 3). Their measurements correspond to HD110281 with A(F) = 2.71 at [Fe/H] = −1.56, and HD135148 with A(F) = 2.59 at [Fe/H] = −1.96. HD135148 has the highest C abundance in their sample and is a spectroscopic binary with a companion star with the approximate mass of a white dwarf (e.g., Car- ney et al. 2003). Li et al. (2013) also applied 3D abundance corrections on the observed sample using 3D models from the CIFIST atmosphere grid (Ludwig et al. 2009). Since the grid of CIFIST do not have 3D models at low gravities as their star sample, four 3D models with comparable stellar parameters but higher gravity were selected to inspect the trend of the abundance corrections with changing Teff and log g . Li et al. (2013) suggested minor 3D effects for the low- gravity giants in their sample and substantial corrections (> 0.1 dex) for gravities log g > 2.0, at [Fe/H] = −2.0. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | The F abundance in CEMP-s stars is assumed to come from an AGB companion (now an unseen white dwarf) which also synthesized C and s-process elements (Lugaro et al. 2004; Ivans et al. 2005; Lucatello et al. 2005; Karakas & Lattanzio 2007; Thompson et al. 2008; Lugaro et al. 2008). Within this picture, theoretical models have to consider not just AGB stellar nucleosynthesis at low-metallicities, but also the binary evolution. Abate et al. (2015a,b) have undertaken an appropriate theoretical study (see Section 2.1). The initial abundances used in Abate et al. (2015a,b), including fluorine, were adopted from the Galactic Chemical Evolution (CGE) model by Kobayashi et al. (2011a). In this CGE model, the contribution from massive stars is set to minimum and does not include that from WR stars. Although Kobayashi et al. (2011b) have suggested that the F production through neutrino-process can be increased by a factor of 100, this is negligible once mass transfer occurs from AGB stars in the initial-mass range [0.9,2.5] M(cid:12) . |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | HE 1429–0551. Our observed F abundance, A(F) = +3.93, is slightly overpredicted but in reasonable agreement with the theoretical predictions from Abate et al. (2015b), A(F)Abate = +4.40. In addition to our F measurement, observed abundances of C and neutron-capture elements by Aoki et al. (2007) are well reproduced by the model for HE 1429–0551. Therefore, the observed chemical pattern of HE 1429–0551 supports the formation scenario on a binary system through mass transfer from a former AGB companion, even though its binary status has not been observationally confirmed as noted in Section 2.1. The best-fit model from Abate et al. (2015b) for HE 1429–0551 correspond to a very long-period binary with an initial period Pi = 1.37 × 105 days. The initial masses of the primary star (the CEMP progenitor) and secondary star (the CEMP star) are M1,i = 1.30 M(cid:12) and M2,i = 0.71 M(cid:12) , respectively. The adopted PMZ mass of the primary star is MPMZ = 10−3 M(cid:12) and the calculated accreted mass by the secondary stars is Macc = 0.17 M(cid:12) . The current estimated period of the system is Pf = 2.3 × 105 days. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | HE 1305+0007. Our F abundance upper limit, A(F) < +3.28, does not agree with the theoretical prediction from Abate et al. (2015b), A(F)Abate = +4.52 dex. This is 1.24 dex higher than our estimated upper limit in this star (see Figure 2 and 4). Ours and the Lucatello et al. (2011) upper limit in this star, both disagree with the predicted abun- dances. The disagreement is very likely due to the fact that HE 1305+0007 is classed as a CEMP-s/r star, which is a sep- arate class of CEMP stars whose origins may be completely different from CEMP-s (the origin of CEMP r/s from an AGB nucleosynthesis point of view was also discussed by Lugaro et al. 2012). Our observation, as well as literature data from Lucatello et al. (2011), suggest a lower fluorine production by the HE 1305+0007 progenitor that cannot be reproduced by the nucleosynthesis model adopted by Abate et al. (2015b) while simultaneously reproducing the very en- hanced abundances of the neutron-capture elements. As dis- cussed by Abate et al. (2015b), in most CEMP-s/r stars the abundances of heavy neutron-capture elements (e.g. Ba, La, Ce, Pb, but also Nd, Sm, and Eu) are strongly enriched and to reproduce their enhancements, the algorithm determin- ing the best-fitting model selects a solution where a large mass of very enriched material is transferred from the AGB primary star to the main sequence secondary star of the binary system. Consequently, the best-fitting model tends to overestimate the abundances of the light elements (e.g. C, N, F, Na, Mg), which are typically enhanced to a simi- lar extent as ”canonical” CEMP-s stars. This result was in- terpreted as an indication that CEMP-s/r star might have undergone a different nucleosynthesis history compared to CEMP-s stars, the abundances of which are generally well- reproduced by AGB nucleosynthesis models (Lugaro et al. 2012; Abate et al. 2015b). |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | et al. (2011). The modelled abundances by Hampel et al. (2016, 2019) can satisfactorily reproduce the observed abundances of C and s-process elements, as well as the r-process elements present in HE 1305+0007. Currently, F abundances from i-process models are not available in the literature that we are aware of. It would be very interesting to examine and test if those models can also reproduce the observed abundances of fluorine as well as the s/r (or i) chemical pattern. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Five stars in the sample from Lucatello et al. (2011) are in common with the theoretical work of Abate et al. (2015b), including HE 1305+0007 which is also in common with this work. Figure 4 shows the comparison between the predic- tions from Abate et al. (2015b) with the observations from Lucatello et al. (2011) and ours. There are some differences in the stellar parameters adopted by Lucatello et al. (2011) and Abate et al. (2015b) (i.e. Teff and [Fe/H] ). The dif- ferences in the adopted [Fe/H] and the predicted-observed abundances for each star are shown in Figure 4 using a con- necting black dashed line. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | The fluorine abundances observed by Lucatello et al. (2011) include CEMP-s and CEMP-s/r. Their measurement in the CEMP-s HD 5223 is in good agreement with the model. However, all their upper limits are highly overesti- mated by the model predictions from Abate et al. (2015b) (see Figure 4). The inconsistencies with the upper lim- its of the CEMP-s/r stars CS 22948–027, HD 187861 and HE 1305+0007 can be accounted for by the issues of the model on reproducing the s/r pattern in addition to the differences in the adopted stellar parameters mentioned be- fore. Nevertheless, our upper limit in HE 1305+0007 was derived using the same stellar parameters as in Abate et al. (2015b), and we found similar discrepancy of 1.24 dex. Ad- ditionally, the upper limit in the CEMP-s star CS 22942–019 estimated by Lucatello et al. (2011) is highly overestimated by the model. The inconsistency with the model in this star cannot be justified by just the differences in the adopted stellar parameters. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Figure 5. A(F) and A(Ba) abundances for CEMP-s, CEMP-s/r and CEMP-no objects from Lucatello et al. (2011) and our sample. Symbols are the same as Figure 3 (carbon-normal field halo stars are excluded). The black dashed line shows the differences in the adopted A(Ba) and the estimate upper limit in our in common observed star with Lucatello et al. (2011), HE 1305+0007. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Nomoto et al. 2013; Limongi & Chieffi 2018). Measured abundances of Fluorine (or strong upper limits) in CEMP- information on the chemical no stars would give useful origins of CEMP-no stars. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | in CEMP-s/r progenitors. Unfortunately, due to the lack of measurements and the large spread in the only two F upper limits in CEMP-no stars, no valuable information can be extracted from its F estimates. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Very recently, Ryde et al. (2020) compared the current literature data on F abundances at higher metallicities, in addition to their own sample, with O and the s-process ele- ment Ce. They showed a flat trend when studying the [F/Ce] ratio as a function of [Fe/H] for intermediate metallicities, −0.6 < [Fe/H] < 0.0. This result strengthens their conclusion on AGB stars as the dominant source of F in that metal- licity range. However, their sample includes two stars below [Fe/H] < −0.8 exhibiting large s-process element (Zr, La and Ce) enhancements yet surprisingly low F abundances. (see Section 5.6 and Figure 10 from Ryde et al. 2020). An in- spection of radial velocities in these two stars by Ryde et al. (2020) did not show any signs of binarity, weakening the possibility of pollution from an AGB companion. A similar result was found by Cunha et al. (2003) in ω Cen stars at similar metallicities to those from Ryde et al. (2020). This raises multiple questions on how these objects obtained their chemical signature and what is its astronomical origin. More data are needed to observationally quantify any relation be- tween s-process and F nucleosynthesis in the early Galaxy. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Our fluorine measurement in HE 1429–0551, A(F) is in reasonably agreement compared to the = +3.93, predicted abundance from Abate et al. (2015b), A(F)Abate = +4.40 (see Figures 2 and 4). The best-fitting model of Abate et al. (2015b) also reproduces the observed abundances of C and s-process elements determined by Aoki et al. (2007). As noted in Section 2.1, the binary status of this object is inconclusive (Jorissen et al. 2016). Without a constraint on the orbital period, Abate et al. (2015b) found that the observed abundances were best reproduced with a very long-period (+10 000 days) binary mode with an initial period of Pi = 1.37 × 105 days, initial masses M1,i = 1.30 M(cid:12) and M2,i = 0.71 M(cid:12) . The adopted PMZ mass of the progenitor and the calculated accreted mass by HE 1429–0551 that account for the observed neutron-capture elements and F abundance is MPMZ = 10−3 M(cid:12) and Macc = 0.17 M(cid:12) , respectively. This result strongly suggests that HE 1429–0551 is indeed a binary system which has not been monitored for a sufficiently long time to constrain its orbit. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | in HE 1305+0007, A(F) < +3.28, significantly lower than the predicted value from Abate et al. (2015b), A(F)Abate = +4.52 (see Figures 2 and 4). Lucatello et al. (2011) also found a F upper limit in HE 1305+0007 of A(F)Lucatello < +3.5, comparable to ours if we consider the differences in the analysis (see Section 4.1). |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | All the observed upper limits in CEMP-s/r stars from Lucatello et al. (2011) and this work show low F abundances compared to CEMP-s (See figure 4 and 5). These observational results suggest that the CEMP-s/r progenitors might provide lower F abundances. The discrep- ancies between the observed upper limits and the predicted fluorine abundances by Abate et al. (2015b) can be due to the CEMP-s/r category type, and the theoretical approach using just s-process nucleosynthesis. Hampel et al. (2016, 2019) used i-process nucleosynthesis models getting a better agreement with the chemical pattern observed in CEMP-s/r stars, including HE 1305+0007. Although the details on how the operation of the i-process affects the production of F has not yet been fully explored, F abundances can be the key to understand the astronomical site of i-process. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | abundances from literature data. Although caution has to be taken in the interpretation of Figure 5, the data may suggest a positive correlation on Ba and F abundances in CEMP-s stars. Additionally, we note that the only three F measurements from CEMP-s objects show higher F content than those in CEMP-s/r. This could indicate differences in the relation between s-process and F nucleosynthesis in CEMP-s and CEMP-s/r progenitors. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | The fluorine observations presented in this paper, along- side literature data, suggest further tests and constraints on nucleosynthesis models for CEMP star progenitors. More- over, the relation between the production of fluorine and s-process elements in metal-poor AGB stars is still uncer- tain, and thus the true nature of its production mechanism remain unclear at these metallicities. Despite the difficulty of the analysis, it is very important to extend the study to additional CEMP stars, in order to gain new insights upon nucleosynthesis and chemical enrichment in the early galaxy. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | This work used the Immersion Grating Infrared Spectrome- ter (IGRINS) that was developed under a collaboration be- tween the University of Texas at Austin and the Korea As- tronomy and Space Science Institute (KASI) with the finan- cial support of the US National Science Foundation under grants AST-1229522 and AST-1702267, of the McDonald Observatory of the University of Texas at Austin, and of the Korean GMT Project of KASI. These results made use of the Lowell Discovery Telescope at Lowell Observatory, supported by Discovery Communications, Inc., Boston Uni- versity, the University of Maryland, the University of Toledo and Northern Arizona University. Parts of this research were supported by the Australian Research Council Centre of Excellence for All Sky Astrophysics in 3 Dimensions (AS- TRO 3D), through project number CE170100013. We thank the referee for her/his helpful comments that improved the clarity of this paper. CA thanks Richard Stancliffe for many inspiring discussions. CK acknowledges funding from the United Kingdom Science and Technology Facility Council through grant ST/M000958/1 and ST/R000905/1. A.MG. acknowledges the support by CONICYT (Chile) through Programa Nacional de Becas de Doctorado 2014 (CONICYT-PCHA/Doctorado Nacional/2017- 72180413). A.MG. thanks Thomas Nordlander and Melanie Hampel for helpful discussions, and also dedicates this paper to all those who have been fighting for greater equality and better edu- cation in Chile. |
Processed_Constraining_Nucleosynthesis_in_Two_CEMP_Progenito.txt | Data available on request. The data underlying this article will be shared on reasonable request to the corresponding author. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Keywords: anomaly detection motorway flow prediction deep learning CNN LSTM BPNN short vs. long-term prediction. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Congestion prediction represents a major priority for traffic management centres around the world to ensure timely incident response handling. The increasing amounts of generated traf- fic data have been used to train machine learning predictors for traffic, however this is a chal- lenging task due to inter-dependencies of traffic flow both in time and space. Recently, deep learning techniques have shown significant prediction improvements over traditional models, however open questions remain around their applicability, accuracy and parameter tuning. This paper bring two contributions in terms of: 1) applying an outlier detection an anomaly adjust- ment method based on incoming and historical data streams, and 2) proposing an advanced deep learning framework for simultaneously predicting the traffic flow, speed and occupancy on a large number of monitoring stations along a highly circulated motorway in Sydney, Australia, includ- ing exit and entry loop count stations, and over varying training and prediction time horizons. The spatial and temporal features extracted from the 36.34 million data points are used in various deep learning architectures that exploit their spatial structure (convolutional neuronal networks), their temporal dynamics (recurrent neuronal networks), or both through a hybrid spatio-temporal modelling (CNN-LSTM). We show that our deep learning models consistently outperform tradi- tional methods, and we conduct a comparative analysis of the optimal time horizon of historical data required to predict traffic flow at different time points in the future. Lastly, we prove that the anomaly adjustment method bring significant improvements of using deep learning in both time and space. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | The parametric approaches were typically based on time-series analysis, Kalman models, etc. For example, the ARIMA model has been widely applied for traffic flow prediction ([22], [11]) due to its simplicity and good perfor- mance in forecasting linear and stationary time-series. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Further extensions of ARIMA have been proposed to account for seasonal features (SARIMA [13]), and for ad- ditional explanatory variables (ARIMAX [24]). The effectiveness of parametric models can be affected by the traffic stochasticity, and by the occurence of disruptive events. As a result, non-parametric models have seen an increas- ing popularity, and among them we cite: k-nearest neighbours [5], support vector regressions [10], artificial neural networks [12], and Gaussian Processes [9]. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Recently Deep Learning (DL) methods have emerged as popular non-parametric alternative approaches for short- term predictions, with various models being proposed and tested in different set-ups. Two major literature reviews on DL models can be found in [3], and [17]. These debate how different DL models can be adapted for traffic flow prediction, and why the spatial and temporal correlation in traffic congestion propagation makes the application of such models difficult. More recently, recurrent neural networks (RNNs) – such as the long short-term memory model (LSTM) – have been designed to learn from sequences of data, and to capture long-term temporal patterns. LSTM was applied to traffic flow data [21], either alone of jointly with convolutional neuronal networks (CNN) [16, 25] in an attempt to capture the spatial road network information. Among the difficulties of deploying such models are their often complicated structure, the choice of parameters (such as the number of neurons or the non-linear functions), and the fact that neural networks have been long time regarded as “black-boxes” [28]. These difficulties are starting to ease due to the emergence of integrated modelling and fitting frameworks, such as TensorFlow [2] and PyTorch [18]. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | There are several open questions relating to the usage of DL methods for traffic flow prediction, still not addressed in the literature. The first question relates to the scalability of such methods. The majority of existing studies concentrate on one or several stations, or over short periods of time [20]. It is therefore unclear how DL models behave at the level of an entire motorway, or for large datasets with complicated road structures. We address this question by constructing deep learning models capable of predicting the real-time traffic flow along an entire motorway. Our dataset spans over an year, and it spreads across 208 stations along 48kms of road network in Sydney, Australia. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | The second open question relates to the relationship between the training and the prediction horizons. The majority of existing work performs traffic flow prediction for future time horizons of 2-3 time intervals [19] however they usually fix the past horizon to 30 or 40 minutes. This arbitrary choice can affect the prediction performances, and impact the model selection. In our study we propose a sensitivity analysis between various past and future horizons for each of the deep learning models under investigation, and we showcase the best set-up for each model. We find that for LSTM there is a limit in the past horizon beyond which the accuracy stops improving, and that CNN performances are actually hurt by using too much data from the past. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | The third open question is about deploying hybrid deep learning models that combine both spatial and temporal modelling. While some research has shown that hybrid DL architectures can improve performances in specific cir- cumstances [20, 23], other studies still debate the necessity and overhaul of fine-tuning such models. In this paper, we implement a hybrid CNN-LSTM model and we find that it under-performs the LSTM model, with its performances fluctuating significantly with respect to the future time horizon. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | The fourth and last question is regarding the efficiency of such models under outliers and anomalies that can affect their accuracy and application. Anomaly detection methods have been widely used in various domains, but in transportation majority of work is reporting towards the usage of surveillance video cameras [4] or combination of machine learning and video-camera streaming [1]. Other methods have been proposed and developed separately [7], but very few in combination with deep learning algorithms or integrating the traffic behaviour revealed by the fundamental diagram of speed-flow for example. We address this question by proposing an outlier detection methods for all incoming streams in the data set (flow, speed, occupancy) and an anomaly improvement/repairing methodology which adjusts parameters not only based on historical data profiles but also on incoming data streams. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | The objective of this paper is to construct an advanced deep learning framework for predicting the traffic conges- tion, which can be used to perform model comparison under varying prediction conditions including outliers/anomalies and which can serve as a benchmark for future predictive work. We apply our models on data coming from counting stations along a long motorway in Australia, and we consider both the spatial structure of the datasets, as well as the historical records during a one year long dataset. We first present the data profiling, outlier detection and anomaly repair method, followed by the methodology for constructing various deep learning models such as CNN, LSTM and the hybrid architecture CNN-LSTM; we then compare them against typically employed approaches, such as ARIMA, individual station regressors, and the average historical traffic profiles. We follow with a sensitivity analysis of histor- ical versus prediction time horizons, and their impact on model performance. We end with the comparative analysis of the prediction performances of the models under both outlier or regular data streams arriving, and conclude with a model choice discussion. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | We first gather the spatial information with regards to the placement of traffic monitoring stations along the mo- torway segments, the road network geometry file and the temporal information in the form of traffic flow recorded at each time step (3min time-intervals in our case). The obtained dataset is described in Section 3.1. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | This step is necessary for building regular traffic patterns depending on the type of day, time-of-day, etc. Fig. 1 showcases 3 different possible cases of station spatial structure, which are automatically checked for data accuracy and motorway structure consistency in both time and space. Let 𝑇𝑓 (18A) and 𝑇𝑓 (19A) be the traffic flows registered at the stations 18A, and 19A respectively. If no exit or entry is recorded between two consecutive stations (see Fig. 1-𝑏1 ), we assume that the two flow patterns should match, and we check for consistency as 𝑇𝑓 (18A) = 𝑇𝑓 (19A) ± 𝜖 (𝜖 accounts ), the module verifies that the sum of for the inherent detector equipment error). When exit loops exist (see Fig. 1-𝑏2 the flow recorded at the downstream stations (3A and 3X) matches the traffic flow recorded at the closest upstream station: 𝑇𝑓 (4A) = 𝑇𝑓 (3A) + 𝑇𝑓 (3X) ± 𝜖. Lastly, in case of entry loops (see Fig. 1-𝑏3 ), the module checks that the sum of downstream traffic flow is the sum of upstream flows: 𝑇𝑓 (11A) = 𝑇𝑓 (12A) + 𝑇𝑓 (12E) ± 𝜖. The data processing step also builds the daily flow patterns for all the stations along the entire motorway. These are further used in the deep learning methodology, and to identify missing data and abnormal traffic flow due to traffic disruptions. This is further discussed in Section 3 together with speed and occupancy profiling as well. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | the future we want to make the prediction) and ̂𝑋𝑡+𝑃 = station over the prediction horizon. A summary of all the notations used in this paper is provided in Table 1. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Deep learning is usually used to learn high dimensional functions via sequences of semi-affine non-linear trans- formations, and it has been shown effective general function learners [19]. A deep learning predictor is capable of addressing the non-linearity in the datasets, and of finding the spatio-temporal relations between features. We imple- ment various DL models in the current modelling framework and we apply them either individually – Convolutional Neural Networks (CNN), Long Short-Term memory networks (LSTM), back-propagation neuronal networks (BPNN) – or in hybrid structures as an advanced deep learning architecture. We compare their performance to other parametric and baseline models in Section 5.1. In the following we detail each of the DL models and we provide their internal architecture setup. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Back-propagation Neuronal Networks (BPNN): BPNN is a typical feed-forward type network which learns the relation between inputs and outputs without an explicit mapping of the information, and using a gradient descent optimisation method. Such models have been successfully applied for highway traffic incident detection [6], and also for tourist volume forecasting in Baidu [15]. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | The topology of BPNN usually includes an input layer, one or multiple hidden layers, and an output layer. In our DL work, we developed a BPNN model which consists of two fully-connected layers. The input of first layer is the historical information of all stations, and the last layer’s output is the prediction of the traffic flow across all monitoring stations. In this work, BPNN is mainly used as a lower-bound DL performance measure, and it serves to assess the performance gains obtained when implementing the more complex models detailed here below. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | a CNN can include both a spatial dimension and a temporal dimension, and the convolution operator can be applied in either one dimension (1D) or both dimensions (2D). In various advanced DL architectures, CNN models usually contain various convolutional layers with non-linear activation functions applied in between. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | For our study, we construct a fully-connected CNN structure as presented in Fig. 2, which accepts as input the RxN feature matrix 𝑿𝒕 defined in Eq. (1). This 2-dimensional input is passed through two convolutional layers and two Rectified linear activation unit (ReLU) functions, before finally being flattened as a 1-D vector and sent through a Fully Connected Layer which outputs the final results. No pooling layers are employed in our model; although, in general, pooling layers may increase the speed of training and achieve better performance on high dimensional image recognition tasks, the spatial dimension of our data set is smaller and the information it contains is not redundant like in images. For advanced DL architectures, the ReLU function are more popular than the traditional sigmoid/tanh functions because: a) they are more computationally efficient, b) they can help avoid the exploding and vanishing gradient problems and c) they tend to show better performance in practice [14]. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Fig. 3 showcases the structure of the LSTM model that we develop in this work for the traffic flow prediction. An LSTM unit is typically comprised of an input layer, a hidden layer (which acts as a memory block containing an input gate, a forget gate and an output gate) and an output layer. The LSTM model is a sequence of LSTM units, in which the output of one unit is consumed as input by the following unit. The output of the last unit feeds into a fully connected layer which makes the the final prediction. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | In our application, the input feature matrix 𝑿𝒕 is split into various flow vectors, one for each counting station (denoted as ⃗𝑺𝒕−𝒊). For each time step from the past horizon {𝑡 − 𝑅 + 1, ..𝑡 − 1, 𝑡}, an LSTM unit accepts the vector ⃗𝑺𝒕−𝒊 as input and outputs a hidden state vector ℎ𝑖 and an output vector 𝑜𝑖, of equal lengths. The hidden state ℎ𝑖 is passed at the next unit {𝑡 − 𝑖 + 1}. The last output vector 𝑜𝑡 is connected to the fully-connected layer to get the final result. For our work, we only used only one hidden layer per unit and no drop-out layers, as we have a limited number of counting stations. For a more complex road network, this LSTM structure can be further extended. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Hybrid CNN-LSTM prediction: We develop the hybrid model as a combination of spatial and temporal processing, modelled by connecting the output of CNN to the input of each LSTM unit. The intuition is that the structure of LSTM would learn the temporal patterns, while the structure of CNN can learn the location features. The final prediction is made using a Fully Connected layer, just like for LSTM. Several previous research works have employed the hybrid model and they report contradictory results: some argue that it improves the prediction accuracy, while others indicate the contrary [17]. Fig. 4 showcases the structure of our hybrid model; the difference from a regular LSTM is that we use a 1-dimensional filter to scan the input (each ⃗𝑺𝒕−𝒊 is processed by a 1-D convolution layer before it is fed into an LSTM unit). The added convolution layer has a 1x3 filter. The results of each proposed model have been compared to other basic or parametric prediction models further detailed in Section 4. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Our methodology has been applied to a motorway traffic dataset, which was collected over the entire year of 2017, by recording the traffic flow, speed and occupancy at each of the 208 bi-directional “counting stations” along the M7 Motorway in Sydney, Australia (shown in Fig. 1). The M7 motorway runs on the West of Sydney and it is the main motorway connecting North and South Sydney. There are 104 metering stations in each direction including entries and exits; stations ending in A denote south-bound traffic, stations ending in B denote north-bound traffic while stations ending in E and X denote entries and exits respectively. The dataset contains 36.34 million data points for flow only, where for example one data point is the flow recorded by one station during one time-interval of 3min, denoted as TI. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Figure 5: (a) Constructing the daily profile. Mean (solid line) and the 20% − 80% percentiles (red area) for the traffic flow series for the station 50A, computed on the period 2017-02-01 to 04-30; (b) Daily profiles for days of the week. The daily profiles for station 02A for each of the days of the week, computed for the same period of time. (c) Daily profiles for all stations – the Traffic Flow Congestion Map. The colormap of the Monday traffic flow for all 104 south-bound (A) stations is calculated based on a Flow/Capacity ratio and ranges between 0 and 1. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Daily profile. We start from the observation that the traffic flow at any given station presents a strong daily and weekly seasonality, mainly driven by the users daily work commute patterns. We define the daily profile as the typical daily traffic flow series recorded at a given station. We compute a station’s daily profile as the average flow for each TI, for a given day of the week, over a period of three months (12 weeks in the period of February 1𝑠𝑡 to April 30𝑡ℎ 2017). Fig. 5a shows the mean flow for a Tuesday for the station 50A, alongside with the 20𝑡ℎ and 80𝑡ℎ percentile values. We make several observations. First, the 20∕80 confidence interval wraps closely the daily profile, indicating that the 12 series are very similar and that the daily profile is a representative summarization. Second, we observe that the daily profile shows two peaks, corresponding to the two rush hours: one in the morning (8-10 AM) corresponding to the daily commute towards work, and a second one in the afternoon (5-8 PM) corresponding to the end of the working day. Last, the daily profile allows to identify non-standard days; for example, the blue line in Fig. 5a shows a significantly lower traffic, as it corresponds to April 25𝑡ℎ – ANZAC day, a public holiday in Australia. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Weekdays vs. weekend. Fig. 5b plots the daily profiles for each of the days of the week, for station 02A. We observe that the weekdays (Monday to Friday) exhibit similar two peak patterns, whereas Saturday and Sunday have a single peak between 11AM and 4PM, and an overall lower flow. Noteworthy, the ANZAC day flow shown earlier in Fig. 5a resembles a weekend pattern, despite being on a Tuesday. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Traffic flow congestion map. Fig. 5c plots as a congestion map the Monday daily profiles of all South-bound stations by calculating the flow/capacity ratios. Here we also observe the two peak patterns of a typical weekday. However, the traffic flow also allows to visually identify the most congested sections of the motorway (between 31A and 13A during the afternoon peak) and most importantly to track down abnormal congestion disruptions along the motorways. In addition to providing an abstraction of the typical flow, the daily profile is also used to correct the missing data in the dataset, most often occurring due to malfunctioning traffic recording devices for particular stations. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Figure 6: Speed daily profile for stations a) 03𝑋 and b) 04𝐴 respectively. Occupancy daily profiles for stations c) 03𝑋 and d) 04𝐴 respectively. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | follow very similar trend at lower speed records, which has a significant speed drop during all morning peak periods between 06 ∶ 30𝐴𝑀 and 09 ∶ 30𝐴𝑀 to almost 55[𝑘𝑚∕ℎ]; this can reflect congestion for exiting the motorway into sub-arterial roads. Similarly, the ODPPs for 03𝑋 reflect higher peak during the same time as the speed drops in the morning rush hours of various weekdays, while for 04𝐴 the ODPPs score higher values for week days profiles, but very closely related to the SDPPs of the same stations. All DPPs for speed, occupancy and flow have been used for: baseline comparison of proposed DL models and anomaly detection and repairing procedure detailed in the following sections. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Figure 7: The total number of missing data points in the Sydney Motorway traffic flow dataset, aggregated per month. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Upon initial inspection, the Sydney Motorway traffic dataset contains missing data points from the real data stream- ing which are misaligned with the overall traffic behavior in that area. We detect such malfunctions at the level of the entire dataset, and we count them per month. Fig. 7 shows that there is an abnormally high number of missing data in the months of May and December 2017. In our initial work, we excluded these two months from our training data (as discussed in Section 4 of the main paper) and conducted our Deep Learning Training on a selected months form our data set. However, a real-time prediction engine should be applied on all type of data sets, and should be able to detect very fast anomalies and outliers that might be transmitted in real time. For this reason, we have extended our previous work and constructed an anomaly detection engine which follows several steps described in the next subsec- tion. The final results will focus on presenting results under both types of set-ups: those using only the original data set without any anomaly detection and those using the outlier processing procedure making use of daily profiles and data interpolation presented in the following. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Before developing the Deep Learning Methodology and its relevant models deployed for our study, we firstly apply an anomaly detection and data repairing procedure which is represented in Fig. 8. Based on the incoming data sets such as flow, speed and occupancy of the motorway, we apply several steps to: a)identify the outliers in the data (these can be of different types and can appear as either missing records, all-zero records or extremely large records), b) to treat the outliers by applying various steps such as interpolation, adjacent data completion based on motorway structure and fundamental diagram represenation of each station, and finally c) annotate the repaired data set and prepare the release for deep learning model training, testing and validation. We make the observation that once the DL models have produced the predicted results over a future selected time horizon, the procedure can repeat again based on new incoming data sets for each motorway station. The following sub-sections detail all steps and modules presented in Fig. 8 by looking into the motivation of addressing each problem and the description of steps taken to address them. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | is not complete, some time points might be missing from being transmitted. Such missing records may be the result of equipment maintenance or can arise from the network failure. Without addressing these missing points, the data set will not be continuous and hard to be used to train any deep learning models. Fig. 9 showcases an example of missing records from an entire 24-hour traffic flow profiling; the Ox axis represents the time interval (with a 3-min frequency) while the Oy axis represents the listing of traffic flow for each station from 02𝐴 until 72𝐴. The traffic flow is coloured by its values with highest values reaching 5,000 vehicle/3-min time interval. The dark region represents almost 3 hours of missing data points from the stream which have unidentified cause. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Figure 9: Example of missing records. The records in black area are not included in original dataset and may contain valuable information on how the congestion evolved during the missing time-period. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Motivation: The initial data profiling showed for example that in some cases, a station’s flow, speed and occupancy can suddenly drop to zero while its neighbourhood stations (preceding and following) continue to report high traffic flow counts for example. When no entries/exits are present in between consecutive traffic flow stations, then the data profiling records should all follow similar trends. If one of the consecutive stations does not align with its neighbours reporting data streams, this indicates another erroneous data transmission or device malfunction. Similar to the missing records, we also want to locate all abnormal all- zeros records that might occur and repair/adjust them in following steps. Fig. 10a showcases an example of all-zeros records recorded by the station 23𝐴 which should follow the same traffic flow trends as its neighbours 22𝐴 and 24𝐴 as no entries/exit ramps are in between them. However the flow recorded by station "23𝐴" seems to be suffering from various drop-outs throughout the day possibly due to device malfunctioning or wrong-reporting. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Figure 10: a) Example of abnormal all-zeros records. This figure shows the flow curve of three adjacent stations in same day. While the flows of previous station (24A) and next station (22A) are not changed significantly, the flow recorded by station 23A suddenly drop to zeros in several times. b) Example of high-flow records. This figure shows the flow curve of three adjacent stations in same day. While the flows of previous station (14A) and next station (16A) are as usual, the flow recorded by station 15A suddenly drop to zeros and last for a while. Then the flow sharply increases to a very high value in next time point. |
Processed_Traffic_congestion_anomaly_detection_and_predictio.txt | Motivation: Another finding of the data profiling is that a few traffic count stations can present continuous abnormal all-zeros records which can last up to several days not only a few hours as described in the first step of this subsection. Such long abnormal all-zeros periods will heavily affect the correctness of the daily profiles and also the training of any deep learning models that we deploy. So we need to detect and repair this type of anomaly before conducting other data preprocessing or analytic investigations. |
Processed_Magnesium_Isotopes_in_Metal-Poor_Dwarfs,_the_Rise_.txt | We have determined magnesium isotopic ratios (25,26Mg/Mg) in metal-poor (−2.6 ≤ [Fe/H] ≤ −1.3) halo dwarfs employing high S/N (90-280) high spectral resolution (R = 105) Keck HIRES spectra. Unlike previous claims of an important contribution from intermediate-mass AGB stars at low metallicities, we find that the rise of the AGB contribution in the Galactic halo did not occur until intermediate metallicities ([Fe/H] & −1.5). |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.