Molecular Simulation Meets Machine Learning

Abstract

Molecular simulation that encompasses both Monte Carlo and molecular dynamics methods, coupled with ever-increasing computing power, has provided very valuable insights linking the nature of intermolecular interactions directly to the macroscopic properties of materials. In contrast, machine learning can be used to predict molecular properties by finding patterns in data rather than directly evaluating molecular interactions. Suitable machine learning approaches for molecules include supervised, unsupervised, reinforcement, and deep learning, with the latter commonly using neural net algorithms. There is considerable overlap in the scope of application of the two approaches, which can be combined for maximum benefit. Careful integration of machine learning with molecular simulation also means that the hypothesis-centered approach of the latter can be both maintained and enhanced. Using machine learning with molecular simulation offers gains in computational efficiency, predictive capabilities, and generalizability. However, the black-box nature of machine learning provides challenges of interpretability and transparency. Data quality, generalizability, and peer review are also issues that require attention. Nonetheless, the combination of the two approaches offers the possibility of greatly expanded predictive capabilities.

The Potential of Neural Network Potentials

“…The most promising tool for doing this is neural network potentials (NNPs), which are a class of machine learning interatomic potentials (MLIPs). These tools are rapidly gaining importance and recognition due to their flexibility and efficacy. − …”