Thesis title: Designing a Physics-Inspired Neural Network Potential for Complex Systems: Application to Water
The simulation of atoms and molecules using computational methods has significantly advanced our understanding of the microscopic world, enabling insights into complex interactions that govern the structure and dynamics of matter. This capability has transformed fields such as materials science and biochemistry, particularly in drug design, by predicting properties of materials and fostering collaboration between theoretical and experimental scientists. Central to these simulations are force fields, which encapsulate physical interactions and allow predictions of molecular behavior. However, the formulation of force fields is complex and system-specific, requiring a balance between microscopic accuracy and macroscopic empirical data. Recent advancements in machine learning, particularly through neural networks, have revolutionized force field development by accurately capturing many-body effects and enabling quantum-level precision in simulations. This shift enhances predictive capabilities and expands the applicability of computational physics to previously intractable systems. My PhD research focused on developing a novel physics-inspired neural network potential (NNP) for both single- and multi-component systems. By implementing the NNP from scratch, we gained deeper insights into the methodology and developed a new training technique that improves accuracy and efficiency. Using water as a case study, we conducted a thorough evaluation of the NNP’s reliability and explored high-pressure states of water, revealing critical physical aspects that warrant further investigation.