Molecular Dynamics (MD) simulations must solve Newton’s equations at the femtosecond scale to resolve atomistic vibrations. However, most phenomena of scientific interest occur at the micro to millisecond scale. This massive timescale discrepancy creates a severe computational bottleneck, requiring traditional MD to run an impractical number of simulation steps. To bypass this limitation, researchers frequently construct reduced-order models by coarse-graining (CG) the atomistic system, enabling larger timesteps. Unfortunately, this computational efficiency comes at the expense of fine-grained atomistic details crucial for many technological applications. While machine learning has accelerated progress by enabling data-driven reduced-order models and generative models to reconstruct fine-grained detail from coarse-grained simulations, ensuring physical accuracy of these models remains a central challenge in the field.
In this talk, we present our recent work toward developing thermodynamically consistent reduced-order models under the GENERIC framework. These models mathematically guarantee that the data-driven model obeys the first and second laws of thermodynamics. Additionally, we introduce our framework for training more physically consistent score-based generative models for the task of recovering fine-grained atomistic details from coarse-grained simulations.