Parametric Reduced-Order Modeling of CH Plasma Mixing Using tLaSDI-pGFINNs

We present a physics-informed, parametric machine learning framework for accelerating molecular dynamics (MD) simulations of mixing in CH plasmas with explicit electrons. Using the Sarkas MD code with a quantum statistical potential, we simulate two-species plasmas across a range of temperatures and extract 1D number and velocity densities for carbon and hydrogen. These densities are Fourier-transformed into 400 real-valued features (50 modes × 2 quantities × 2 species, with complex modes separated into real and imaginary parts) and used as input to a reduced-order model.

Our framework leverages the tLaSDI-pGFINN architecture: a low-dimensional autoencoder representation combined with neural-network-based latent dynamics (LNN and MNN), trained to evolve Fourier modes over time. We train on MD trajectories at 5, 10, and 15 eV, and test at intermediate temperatures (7.5 and 12.5 eV), achieving interpolation in parameter space. While current mean squared error (MSE) metrics show degradation in long-time latent space prediction, ongoing improvements aim to resolve these issues. Once validated, this method is expected to yield speedups of 100×–1000× over direct MD by bypassing costly particle-based simulations. The predicted number and velocity densities can be sampled to reconstruct ion distributions, enabling fast evaluation of transport and mixing metrics in high energy density plasma systems.