Fourier–Thermodynamic Latent Modeling for Temperature-Dependent Plasma Mixing

We present FT-LaSDI (Fourier Thermodynamics-based Latent Space Dynamics Identification), a three-stage reduced-order modeling (ROM) framework for temperature-dependent mixing in quantum-statistical molecular dynamics (MD) of carbon--hydrogen plasmas. Using Sarkas with the Deutsch quantum statistical potential, we train on selected temperatures and evaluate both training-rollout fidelity and interpolation. On training temperatures, forecasting errors remain below 1\% for the first half of the trajectory and rise to approximately 10\% by the final timestep; interpolation at 7.5 eV is not yet successful (relative error $>$50\%). Inference with FT-LaSDI yields an estimated speed-up of 500$\times$ over full MD per trajectory, with a one-time training cost of 24 GPU-hours. This study highlights the promise of thermodynamically consistent ROMs for MD-driven plasma dynamics and pinpoints loss weighting, optimizer choice, and batch size as key levers for robust interpolation.