Accelerating Theoretical Anharmonic Vibrational Analyses with Machine Learning

Solids are fundamental to various industries, including construction, manufacturing, electronics, and transportation, significantly contributing to economic growth and societal advancement. The desired bulk properties of these solids originate at the atomic level, where interatomic forces govern the structural arrangements of atoms and molecules, which in turn determine macroscopic characteristics. Understanding the intricate interatomic forces within solids is notably more complex than in gases or liquids due to their fixed and closely packed structures. This complexity highlights the need for advanced experimental techniques and complementary theoretical frameworks to accurately model these forces. Experimental methods such as vibrational spectroscopy --- specifically Raman and infrared (IR) spectroscopy --- provide critical insights into the forces within solids. However, these techniques often encounter limitations due to factors like time constraints and instrumentation costs. Computational simulations serve as a complementary approach, offering theoretical frameworks to model and predict solid behavior at the atomic level. Traditional methods such as solid-state density functional theory (ss-DFT) have proven valuable for understanding interatomic forces and modeling a diverse set of material properties, including gas adsorption and vibrational dynamics. Conventional approaches to modeling vibrational dynamics frequently employ the harmonic approximation, which, while offering valuable insights, may not accurately represent the system under all circumstances, potentially limiting the model's reliability and applicability. This work illustrates the power of implementing anharmonicity in vibrational simulations while acknowledging the significant challenges due to the need for extensive sampling of the potential energy surface (PES) and high-level quantum mechanical calculations that can be computationally prohibitive for large systems. Specifically, the generation of a PES of sufficient quality remains a major barrier due to the calculation of higher-order derivatives required for accurate modeling. To address these challenges, this research proposes integrating machine learning (ML) methodologies to enhance the efficiency and match the accuracy of traditional anharmonic analyses. By leveraging graph convolutional neural networks (GCNNs), this work aims to develop innovative theoretical frameworks capable of effectively calculating higher-order derivatives of the PES. The findings indicate that ML methods can replicate the harmonic analysis and higher-order derivatives of the PES, providing a promising avenue for future research in solid-state physics and materials science.