Active learning graph neural networks for partial charge prediction of metal-organic frameworks via dropout Monte Carlo

Abstract

Metal-organic frameworks (MOF) are an attractive class of porous materials due to their immense design space, allowing for application-tailored properties. Properties of interest, such as gas sorption, can be predicted in silico with molecular mechanics simulations. However, the accuracy is limited by the available empirical force field and partial charge estimation scheme. In this work, we train a graph neural network for partial charge prediction via active learning based on Dropout Monte Carlo. We show that active learning significantly reduces the required amount of labeled MOFs to reach a target accuracy. The obtained model generalizes well to different distributions of MOFs and Zeolites. In addition, the uncertainty predictions of Dropout Monte Carlo enable reliable estimation of the mean absolute error for unseen MOFs. This work paves the way towards accurate molecular modeling of MOFs via next-generation potentials with machine learning predicted partial charges, supporting in-silico material design.

PACMAN: A Robust Partial Atomic Charge Predicter for Nanoporous Materials Based on Crystal Graph Convolution Networks

“…Very recently, the dropout Monte Carlo method was reported to predict partial charges in MOFs with an active learning graph neural network, achieving an optimal MAE of 0.0083e on the test set with less than 100 atoms in the QMOF database. 86 Our work did not exclude any structures from the QMOF database and still achieved higher accuracy on the test set.…”

“…This indicates that the PACMAN framework is highly effective in capturing and learning the local atomic environment around an atom to accurately predict partial atomic charges, underscoring the model framework’s utility in accurately predicting the partial atomic charge and beyond. Very recently, the dropout Monte Carlo method was reported to predict partial charges in MOFs with an active learning graph neural network, achieving an optimal MAE of 0.0083e on the test set with less than 100 atoms in the QMOF database . Our work did not exclude any structures from the QMOF database and still achieved higher accuracy on the test set.…”