Molecular dynamics (MD) is a powerful computational method used to simulate the motion of atoms and molecules. MD simulations compute the evolution of a system of interacting particles by applying Newton’s equations of motion, facilitating the study of a range of physical, chemical, and biological phenomena. While quantum mechanical (QM) simulations result in accurate predictions of geometries and energies essential for studying various phenomena, the computational complexity has led to the emergence of new approaches such as classical force fields, reactive force fields, and machine learning potentials (MLPs), each offering unique trade-offs. Classical force fields offer longer simulation times due to assumptions such as static bonds and charges, which prohibit the study of reactive systems. Reactive force fields, such as ReaxFF, bridge the gap between QM methods and classical force fields by allowing dynamic bonds and charges. The improved flexibility results in a higher computational load and a more complex functional form that is hand-crafted by domain experts. MLPs are a more recent approach that utilize large datasets to eliminate complex functional forms, while also leveraging the vast ecosystem of machine learning frameworks for enhanced computational efficiency and ease of development.As the number of methodologies increases, the landscape of MD methods becomes more complex, with each method bringing unique attributes and challenges in simulating molecular systems. We introduce innovative hybridization techniques aiming to leverage the strengths of multiple modeling approaches, improving predictive capabilities and computational efficiency. We introduce a hybrid modeling approach called ReaxFF/AMBER that combines the reactivity and polarization capabilities of ReaxFF with the efficiency of classical force fields, facilitating the simulation of larger reactive regions. Although ReaxFF can offer high fidelity when trained carefully, the existing parameterization tools lack the efficiency and speed essential for creating new ReaxFF parameter sets for different applications of interest. We have proposed a novel parameter optimization approach, JAX-ReaxFF, leveraging the capabilities of a scalable machine learning framework to drastically reduce the training times for ReaxFF, thus enhancing the development of new force fields for various applications. We have also modified JAX-ReaxFF to run end-to-end differentiable simulations on different architectures such as CPUs, GPUs, or TPUs with the help of JAX. JAX is a library known for high-performance numerical computing and it provides features such as automatic differentiation and optimization of Python functions. This approach also allows for improved integration with existing machine learning software infrastructure, offering enhanced flexibility and performance portability.Lastly, we propose and compare various uncertainty quantification (UQ) methods suitable for MLPs. These methods are essential for active learning-based data generation approaches, which are crucial for training data-intensive machine learning models. While our primary focus is on MLPs, the datasets created using active learning methods could also enhance the parameterization efforts for classical and reactive force fields.
Read
