Predicting potential energy surfaces with machine learning

M. Hellström
Software for Chemistry & Materials BV,

Keywords: neural networks, molecular dynamics, materials modeling


Atomistic simulations of molecules and materials require a reliable way of evaluating the underlying potential energy surface. Realistic simulations of materials with complexities such as grain boundaries, vacancies, impurities, and interfaces, can only be carried out if many thousands or even millions of atoms are modeled. The sheer size of such systems necessitates the use of a computationally efficient method. Machine learning (ML) methods can bridge the gap between electronic structure calculations - which are accurate but prohibitively computationally expensive - and force fields, - which are fast to evaluate but often not accurate enough. In this work, we describe how an ML method based on artificial neural networks can be applied to describe chemical reactions in several complex systems, such as electrolyte solutions [1] and at solid/liquid interfaces [2]. These types of simulations will give unprecedented detail into the working and degradation mechanisms of batteries and fuel cells, among others. Although neural networks are ideally suited for describing complex non-linear functions like potential energy surfaces, they are still not routinely employed for this purpose [3]. Here, we will present our work for overcoming the main obstacles for widespread adoption, relating to (i) construction of suitable training and validation sets, (ii) automation of the featurization of different molecules and materials, and (iii) the choice of loss function for the neural network optimization. The neural network method is implemented into a software package containing a sophisticated molecular dynamics engine and first-principles, semi-empirical, and atomistic potential methods. This software environment allows for seamless transitions between the different levels of theory and greatly simplifies the construction of the neural network potential. Our ultimate goal is to provide the chemistry, biochemistry, and materials science communities with an all-purpose computationally inexpensive method that can be used in simulations related to solving challenging problems in energy, renewables, climate change reduction and health care. 1. Hellström, M., & Behler, J. (2016). Concentration-Dependent Proton Transfer Mechanisms in Aqueous NaOH Solutions: From Acceptor-Driven to Donor-Driven and Back. The Journal of Physical Chemistry Letters, 7, 3302-3306. 2. Hellström, M., Quaranta, V., & Behler, J. (2019). One-dimensional vs. two-dimensional proton transport processes at solid–liquid zinc-oxide–water interfaces. Chemical Science 10, 1232-1243. 3. Hellström, M., & Behler, J. (2018). Neural Network Potentials in Materials Modeling. In Handbook of Materials Modeling: Methods: Theory and Modeling, 1-20, Springer: Cham