Computational study demonstrates improved dynamical fidelity in molecular simulations, highlighting top-down learning of non-Markovian memory kernels.
Key Points
To develop a differentiable simulation framework that learns non-Markovian memory kernels for generalized Langevin equation closures without requiring explicit projected-force reconstruction.
Combined differentiable simulation with a trainable colored-noise filter to generate stationary random forces and define friction memory satisfying fluctuation-dissipation consistency.
Optimized memory parameters via end-to-end backpropagation through coarse-grained trajectories to match target velocity autocorrelation functions directly.
Tested performance on benchmark models including bulk water, bulk carbon dioxide, and a single-particle star polymer.
DiffGLE enhanced dynamical correlation agreement across bulk water, bulk carbon dioxide, and star-polymer benchmarks compared to standard coarse-grained methods.
The framework accurately reproduced non-Markovian dynamics while fully preserving the underlying conservative potential and structural equilibrium properties.