GNoME aims to expand the catalog of known stable crystals, building upon previous work such as the Materials Project and the OQMD datasets. Snapshots of these datasets, taken at fixed points in time, are used for reproducibility. The GNoME discoveries are based on structures from the Materials Project as of March 2021 and OQMD as of June 2021. By leveraging data from these sources, GNoME aims to enhance the collection of stable crystals, potentially leading to new discoveries beyond the existing datasets.
Structural substitution patterns in GNoME are based on probabilities derived from existing data. The model calculates the probability of ionic species substitution within crystal structures, which aids in discovering new compounds with limited computational resources. GNoME modifies the probabilistic model to increase the number of candidates, focusing on unique discoveries. By adjusting model parameters and thresholds, GNoME promotes novel discoveries and effectively explores composition space. This modification allows a balance between the original model and considering all possible ionic substitutions, preventing combinatorial blow-ups.
To enhance structure generation diversity, GNoME introduces a framework called Symmetry Aware Partial Substitutions (SAPS). This framework enables the efficient discovery of structures with partial replacements, leading to the discovery of new prototypes and structures. GNoME uses SAPS to increase the diversity of structures generated, making the dataset more comprehensive and aiding in the discovery of new materials.
GNoME leverages active learning to generate and evaluate candidate materials through DFT methods. The pipeline involves multiple stages of generating candidates, filtering them using machine-learning models, and evaluating them with DFT. Active learning, combined with GNoME models and dataset expansions through stable crystal structures, has led to the discovery of a substantial number of stable crystals. This iterative process has resulted in over 381,000 new stable crystal structures.
Machine Learning Interatomic Potentials (MLIPs) are trained using a NequIP potential on the GNoME dataset to predict energies and forces. These potentials are used for predicting the dynamic behavior of materials through molecular dynamics simulations. The robustness of these models is tested by training on selected materials and evaluating their performance under different conditions, ensuring reliable predictions for various scenarios.