Accelerating transient structural dynamics via SPILS-Net, a physics-derived latent space subdomain surrogate

Complex, high-fidelity, transient numerical simulations often become computationally prohibitive for tasks requiring thousands of iterations, such as the rapid design space exploration necessitated by automotive crashworthiness assessments. Surrogate models have been widely adopted to overcome this limitation. However, traditional surrogate models and standard Recurrent Neural Networks (RNNs) struggle with complex history-dependent behaviors (e.g., elastoplasticity) due to uninterpretable latent states and the severe training costs of Backpropagation Through Time (BPTT). To address this, we propose the Spatiotemporal Physics-derived Internal Latent Space Network (SPILS-Net), a gray-box surrogate modeling framework. SPILS-Net explicitly models physical history using a physics-derived internal latent space, demonstrated herein via a PCA-reduced latent subspace of accumulated plastic strain. By employing a permanent “Teacher Forcing” strategy, the architecture shifts from sequence-based to increment-based learning. This completely bypasses BPTT and mitigates vanishing gradient problems, reconciling feed-forward training efficiency with the recursive depth required for stable, autonomous inference. Integrated into non-iterative Dirichlet–Neumann coupling workflows, SPILS-Net efficiently replaces expensive high-fidelity subdomains by directly predicting converged interface forces. Evaluated on a complex structural mechanics benchmark, the architecture achieved an 89% reduction in training time, dropping from 7.3 h for a baseline LSTM to 0.8 h, while simultaneously delivering superior predictive accuracy. By decoupling temporal sequence length from training cost, SPILS-Net provides a highly extensible, interpretable surrogate solution. Its architecture scales across extended time horizons, finer mesh resolutions, and increasingly complex physics, effectively unlocking the potential for extensive parameter studies in transient simulations.