Advancing Neural Network Potentials for the Temperature-Dependent Dynamics of Complex Energy Materials

The accurate prediction of temperature-dependent properties in energy materials is critical for the design and optimization of next-generation technologies in energy storage, conversion, and harvesting. However, the state-of-the-art atomistic simulation methods, such as those based on density functional theory (DFT), are computationally prohibitive in predicting these properties at relevant length and time scales due to poor scaling with system size. Artificial neural networks can efficiently approximate the complex potential energy surfaces of these materials, thanks to their expressiveness and computational efficiency, making them the preferred method for accurately simulating temperature-dependent properties at relevant scales. In this talk, I will present recent advances in the development of neural network potentials (NNPs) for capturing temperature-dependent properties in complex energy materials. I will highlight our development of NNPs using accurate DFT data and their application to systems such as elemental carbon, carbon capture in amine-appended metal-organic frameworks (MOFs), and moiré superlattices of stacked transition metal dichalcogenide (TMD) monolayers. Additionally, I will discuss strategies to incorporate electrostatic interactions into NNPs, which are essential for accurately modeling charged systems and interfaces. Furthermore, I will explore our analysis of Behler-Parrinello symmetry functions and their generalization to multispecies systems through machine learning-based species representations, demonstrating clear advantages over computationally expensive equivariant message-passing models for large-scale simulations. These advances can significantly enhance the scalability and accuracy of atomistic simulations, enabling deeper insights into the temperature-dependent properties of energy materials across a wide range of applications.