Development of Mediator Assisted N' GNN Optimization (MANGO) Project

In this work, we developed a methodology to mediate a series of Machine Learning Interatomic Potentials (MLIP). In general, active learning protocols for each MLIP still typically relies on a direct sampling to DFT calculations, resulting in an effectively "low bandwidth" learning process due to the inherent limitation in the DFT sampling. We propose to use a mediation process that interactively connects several known MLIP's, culminating in the development of GNN-based models, so that the "ground truth" sampling needs not to solely rely on the extensive ab-initio calculations. Rather, our Mediator Assisted N' GNN Optimization (MANGO) project aims at creating an ecosystem within which the higher fidelity MLIP's (e.g., the ones employing symmetry equivariant features) can be employed to assist the active learning protocol to optimize the lower fidelity MLIP's (e.g. the ones relying on symmetry invariant latent space). This "high bandwidth" learning approach would allow us to use larger sized samples to help evaluate, for example, long-range interactions and to produce more diverse but relatively higher fidelity MLIP models requiring less computational (GPU) resources. The use of open-source DFT codes would also be integrated to widen the development of MLIP's "for the masses." The partial support from the NASA-MOSGC is gratefully acknowledged.