Molecular dynamics (MD) simulation is an invaluable tool in multiple fields. However, traditional classical mechanics-based MD simulations are usually not accurate enough and quantum mechanics-based potentials are too costly to evaluate. Hence, a machine learning (ML) based potential, that can obtain quantum mechanics-level accuracy with close to classical mechanics-level cost is highly desirable. One promising way of constructing such potential is using a fingerprinting scheme to fingerprint the atomic environment followed by a neural network to predict its corresponding energy for each atom in a system. One great example is the state-of-art Behler-Parrinello scheme, and it works very well with systems with few types of elements present. Unfortunately, to our knowledge, all fingerprinting schemes out there scale poorly with the number of element types present, because the dimensions of the fingerprint increase rapidly, and they are usually forced to use the less scalable high-dimensional neural network structure that uses different subnet for a different element. This is a significant disadvantage that prevents us from constructing universal ML potential with them, which should in theory work for all possible elements. In this work, we propose Gaussian Multi-pole (GMP), a fingerprinting scheme that is not only as accurate and fast as the current state-of-art Behler-Parrinello scheme but also scalable to systems with multiple types of elements. This is because the number of descriptors used to fingerprint an atom is not related to the number of element types in the systems, and it adapts the much more scalable single neural network structure naturally where all elements share the same neural network. Moreover, it is shown that the quality of environment fingerprints with this scheme is systematically improvable and ML potential based on it is systematically convergible. Using the Gaussian multi-pole scheme, a universal ML-based potential is constructed for the open catalyst project that contains 56 types of elements.

GMP: A universal framework for featurization of atomistic systems

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.