Training Algorithm Matters for the Performance of Neural Network Potential: A Case Study of Adam and the Kalman Filter Optimizers

TL;DR

This study compares the effectiveness of Adam and EKF training algorithms for neural network potentials, revealing EKF's superior transferability and robustness, with performance linked to Fisher information rather than validation error.

Contribution

It introduces the implementation of EKF in TensorFlow for training neural network potentials and compares its performance to Adam using water datasets.

Findings

EKF-trained NNPs are more transferable.

EKF is less sensitive to learning rate variations.

Performance correlates with Fisher information measure.

Abstract

One hidden yet important issue for developing neural network potentials (NNPs) is the choice of training algorithm. Here we compare the performance of two popular training algorithms, the adaptive moment estimation algorithm (Adam) and the Extended Kalman Filter algorithm (EKF), using the Behler-Parrinello neural network (BPNN) and two publicly accessible datasets of liquid water [Proc. Natl. Acad. Sci. U.S.A. 2016, 113, 8368-8373 and Proc. Natl. Acad. Sci. U.S.A. 2019, 116, 1110-1115]. This is achieved by implementing EKF in TensorFlow. It is found that NNPs trained with EKF are more transferable and less sensitive to the value of the learning rate, as compared to Adam. In both cases, error metrics of the validation set do not always serve as a good indicator for the actual performance of NNPs. Instead, we show that their performance correlates well with a Fisher information based similarity measure.