Multi-Fidelity Scale Bridging Between Electronic, Atomistic and Mesoscale Using Reinforcement Learning

This thesis presents a comprehensive study on multi-fidelity scale bridging between electronic, atomistic, and mesoscale modeling for silica, utilizing reinforcement learning (RL) to address longstanding challenges in accurately predicting the structural, energetic, alongside other properties of this ubiquitous material. Silica, with its rich polymorphism and diverse applications in fields such as catalysis, water purification, glass production, and energy storage, has proven difficult to model due to the diversity of its Si-O bonding and the wide range of crystal and amorphous phases it can form. Existing models range from computationally efficient empirical potentials, such as the Beest-Kramer-van Santen (BKS) potential, to more recent machine-learned potentials (e.g., GAP, NNPScan), which offer greater accuracy but at a significantly higher computational cost. Despite decades of research, developing a model that balances accuracy with computational efficiency across the full range of silica polymorphs remains an open challenge. In this work, we introduce a novel approach that leverages RL to optimize empirical and machine-learned interatomic potentials for silica, significantly enhancing their accuracy and efficiency. Using multi-reward RL, we systematically explore the high-dimensional parameter spaces of established models, such as the BKS and Soules potentials, to derive new parameterizations that improve the prediction of key properties across various silica phases. For example, our re-parameterized BKS model (ML-BKS) captures not only the equilibrium properties of quartz but also the structure and properties of over 20 metastable silica polymorphs, with a focus on improving the model's flexibility in representing both crystalline and amorphous phases. Additionally, we introduce a new machine-learned potential (ML-Soules), optimized for computational efficiency, that performs on par with state-of-the-art models while requiring fewer computational resources. This thesis also explores three-body interactions, such as those incorporated in the Tersoff potential, to improve the descriptive power of silica models. These interactions allow for better representation of angular dependencies, which are crucial for accurately capturing the mechanical properties and phase transitions of silica polymorphs. The inclusion of three-body terms in the ML-BKS and ML-Soules models enhances their ability to capture the intricate behavior of silica, particularly in non-equilibrium states and high-pressure phases. We demonstrate that such enhancements are critical for improving the predictive accuracy of models across a wide range of polymorphs, including zeolites and amorphous structures. The thesis also introduces a machine learning force-field for silica using MACE framework which is a significant advancement in materials modeling. Unlike traditional models that rely on fixed functional forms, MACE incorporates both short-range and long-range interactions without predefined limitations, making it highly flexible. This approach enables MACE to achieve high accuracy in predicting silica’s structural and energetic landscape while maintaining computational efficiency, thus making it suitable for large-scale simulations and real-time applications like molecular dynamics. The MACE models introduced here represent a paradigm shift in silica modeling, moving away from the limitations of empirical potentials and toward a more flexible, data-driven approach that captures the full complexity of silica’s structural and energetic behavior. By integrating both two-body and three-body interactions, and by leveraging machine learning to optimize parameters, these models provide a more accurate and computationally feasible method for studying silica polymorphs under a wide range of conditions. Additionally, this reinforcement learning workflow is utilized to tune a Coarse-Grained (CG) model for silica, a critical step in developing models capable of simulating large length/time scales. This work lays the groundwork for utilizing high-quality data alongside a state-of-the-art optimization strategy to create coarse-grained models for silica. These models employ simple functional forms while incorporating improved angular resolution, marking an important advancement in the models for silica. In conclusion, this thesis demonstrates the power of multi-fidelity modeling combined with machine learning and RL to significantly improve the accuracy and computational efficiency of silica models. Our work not only advances the understanding of silica’s complex structural landscape but also provides a foundation for the design of novel force-fields or improvements to the exisiting functional forms. Through the development of enhanced interatomic potentials that bridge scales from electronic to mesoscale, this research opens new avenues for the accelerated discovery and synthesis of silica-based materials in a variety of applications.