Spaces:
Running
Apply for community grant: Academic project (gpu and storage)
We are building an open-source, community-inclusive platform to benchmark foundation machine learning interatomic potentials (MLIPs) beyond energy and force error metrics. MLIPs have become increasingly important in both scientific research and industrial applications, as they enable the modeling of complex atomic interactions with high efficiency and accuracy, far surpassing traditional methods in terms of scalability and flexibility. Their ability to predict material properties has vast implications for fields like materials design, chemistry, and drug discovery.
Our ongoing work with Matbench Discovery (MBD) is gaining significant attention as it addresses a critical challenge: testing MLIPs' capability for crystal stability classification.
The best-performing MLIPs today often come from industry and are not publicly available, limiting broader access to state-of-the-art technology. This highlights the essential need to build a community that develops MLIPs for the public good, ensuring that cutting-edge potentials are accessible for academic and general use.
To advance fairness, transparency, and real-world applicability, we propose the MLIP Arena, a broader evaluation framework. This platform will assess MLIPs' quality in various real-world settings, ensuring their robustness and generalization capabilities across different material systems and use cases. Users will be able to contribute their models and benchmark tasks, with an eventual one-click feature to upload simulation scripts and see real-time performance metrics, fostering innovation and community-driven progress.
On Sep 29: Thanks @pbenner for the edit !! 🔥