Abstract
Molecular dynamics simulations are essential for studying complex molecular systems, but their high computational cost limits scalability. Coarse-grained (CG) models reduce this cost by simplifying the system, yet traditional approaches often fail to maintain dynamic consistency, compromising their reliability in kinetics-driven processes. Here, we introduce an adversarial training framework that aligns CG trajectory ensembles with all-atom (AA) reference dynamics, ensuring both thermodynamic and kinetic fidelity. Our method adapts the adversarial learning paradigm, combining a physics-based generator with a neural network discriminator that differentiates between AA and CG trajectories. By adversarially optimizing CG parameters, our approach eliminates the need for predefined kinetic features. Applied to liquid water, it accurately reproduces radial and angular distribution functions as well as dynamical mean squared displacement, even extrapolating long-timescale dynamics from short training trajectories. This framework offers a new approach for bottom-up CG modeling, offering a systematic and principled way to preserve dynamic consistency in complex coarse-grained molecular systems.