Scientific Achievement
Researchers in the New Cooperative Adsorbents and Regeneration Methods Program developed and implemented a machine learning protocol leveraging equivariant neural networks for accurate molecular crystal structure prediction.
Significance and Impact
Opens doors for accelerated materials discovery by significantly reducing the computational cost and time associated with traditional plane-wave based crystal structure prediction methods while achieving accuracy comparable to high-level quantum chemical methods (hybrid DFT, MP2, CCSD(T)).
Research Details
- Created a data curation pipeline for generating representative molecular cluster datasets
- Trained equivariant neural network models efficiently on molecular cluster data, demonstrating their transferability to predict properties of molecular crystal structures
- Integrated the trained models with crystal structure prediction software (e.g., USPEX) to accelerate the search for low-energy crystal structures
Publication Details
A. K. Gupta, M. M. Stulajter, Y. Shaidu, J. B. Neaton, W. A. de Jong, ACS Omega (2024).
