Prof Ong gave an invited seminar talk at the National University of Singapore on Jul 5 2023. In this talk, Prof Ong discusses the different ways in which machine learning (ML) can be used to improve or accelerate the various steps of in silico materials design. The general goal is to preserve the universality and accuracy of ab initio approaches as far as possible while achieving orders of magnitude speed-ups and improved scaling. Prof Ong shared his view that graph deep learning models trained on large diverse materials datasets, such as the M3GNet universal potential, are the “foundation” models for materials science. He further argues that the most robust approach is to replace the smallest, most expensive step in the materials design workflow with ML and preserve as much as the physics of thermodynamics, kinetics, etc. in the computation of materials properties.
We are excited to announce that Materials Graph Library (matgl), our Deep Graph Library/PyTorch reimplementation of the MatErials Graph Network (MEGNet) and Materials 3-body Graph Network (M3GNet) models, is now ready for widespread beta testing! We finally achieved near-feature parity with the original implementations in Tensorflow after months of hard work. The new MatGL includes retrained models of the M3GNet universal potential and the MEGNet formation energy and multi-fidelity band gap models. We have also taken the trouble to include example notebooks to get users started quickly. We believe this new implementation will be more future-proof and extensible. Feedback/issue reports are definitely welcome. This is a collaborative effort between the Materials Virtual Lab and Intel Labs.
Dr Chi Chen’s swansong work in our group, “A Universal Graph Deep Learning Potential for the Periodic Table” is now published in Nature Computational Science! Interatomic potentials (IAPs), which describe the potential energy surface of atoms, are a fundamental input for atomistic simulations. However, existing IAPs are either fitted to narrow chemistries or too inaccurate for general applications. In this work, we combine graph neural networks with traditional 3-body interactions to develop a flexible, yet accurate architecture for machine learning of materials properties. Using the massive database of structural relaxations performed by the Materials Project over the past ten years, we train a universal IAP for 89 elements of the periodic table with broad applications in structural relaxation, dynamic simulations and property prediction of materials across diverse chemical spaces. Using the new capabilities of the M3GNet universal IAP, we are proud to launch matterverse.ai, a ML database of yet-to-be-synthesized materials. Matterverse.ai currently contains about 31 million hypothetical crystal structures, of which about 1.8 million materials were identified to be potentially stable. The database also provides ML properties using state-of-the-art multi-fidelity MEGNet models, such as experimental, HSE and PBE band gaps, bulk and shear moduli, etc. Check out the article here. […]