Learning Dissipative Particle Dynamics without gradient descent

Coarse-graining replaces groups of particles with single effective ones, making time and length scales accessible that fine-grained molecular dynamics simulations cannot reach. The effective interaction, however, is no longer conservative: it acquires a dissipative force and an associated random force by the fluctuation-dissipation relation. Machine-learned forced fields that integrate physical laws offer great promise compared to pure black-box approaches in physically consistent modelling. Optimizing network parameters using iterative gradient descent algorithms (e.g., Adam) is, however, often computationally expensive and can lead to slow convergence. We build the dissipative DPD structure into a random feature graph network instead: the force is pairwise and central by construction, splits into a conservative and a dissipative weight function read out linearly from one sampled feature map. The hidden weights are sampled from the data with Sampling Where It Matters (SWIM) instead of iteratively optimized using gradient descent, so the whole fit reduces to a single least-squares solve. We evaluate our framework on coarse-grained data generated with LAMMPS and report that training the same architecture with gradient descent is $3.9$× worse on accuracy while being $639$× slower. We further evaluate our method and report that it can reproduce the physics of the reference: its temperature, equilibrium structure, and short-time dynamics with very low errors.