Application of PDE emulators for a multi-material, multi-physics simulation problem

Diane Oyen, Morgan Jones, Adithan Kathirgamaraju, Abu Bucker Siddik, Alexander Most, David Schodt, Christine Sweeney, Kyle Hickmann, Michael McCann, Jemma Stachelek, Alexius Wadell, Michal Kucer, Mahindra Rautela, Siddharth Mansingh, James Amarel, Gerd J. Kunde, Emily Casleton Taylor, Soumi De, Ayan Biswas, Aimee Hungerford, Nathan Debardeleben, Earl LawrencearXiv 2026 · DOI · arXiv

This paper from researchers at Los Alamos National Laboratory tackles a critical bottleneck in modern engineering: the massive computational cost of high-fidelity physics simulations. While Machine Learning (ML) emulators—often called "surrogate models"—promise to speed up these simulations by orders of magnitude, they are usually tested on relatively simple benchmarks. This study pushes these models into the "deep end" by evaluating them on a complex, multi-material, multi-physics problem that mirrors real-world industrial challenges.

TL;DR

Researchers evaluated various ML architectures (diffusion models, transformers, and U-Nets) on the challenging Perturbed Layer Interface (PLI) dataset, which involves complex multi-material interactions and high-resolution dynamics. They found that while generative diffusion models performed best, current "foundation models" for PDEs struggle to adapt to the high spatial resolution and long time horizons required for practical engineering applications.

What Problem Does This Paper Address?

Traditional numerical simulators for partial differential equations (PDEs) are accurate but painfully slow, often requiring thousands of CPU cores for a single run. ML emulators aim to replace these with fast neural networks. However, most existing benchmarks (like PDEBench) focus on simple fluid dynamics at low resolutions (e.g., 128x128 pixels) over short time windows.

Real engineering problems, like the one studied here, involve:

  • Multiple Materials: Different substances (copper, aluminum, high explosives, steel) interacting in the same space.
  • Complex Physics: Plastic deformation, phase changes, and material-specific equations of state.
  • High Resolution: The authors use the PLI dataset, which has 27x higher spatial resolution and 5x longer time horizons than standard benchmarks, resulting in data that is over 200 times larger per trajectory.

Methods

The authors compared two main categories of models:

  1. Surrogate Models (Trained from scratch): They used EDM (Elucidating Diffusion Models) in two sizes: EDM-L (64M parameters) and EDM-S (11M parameters). These models were trained to "denoise" the next step in a simulation given the current one.
  2. Foundation Models (Fine-tuned): They adapted SPUS (a convolutional U-Net) and POSEIDON (a transformer-based model) that were pre-trained on simpler PDE datasets.

To evaluate these models fairly, they moved beyond simple error metrics (like RMSE) and used:

  • Pearson Correlation: To measure how well the structure of the physics is captured, especially in sparse regions.
  • Optimal Offset: A technique to account for spatial shifts in predictions, distinguishing between "nearly right but slightly shifted" and "completely wrong structure."
  • Mass Conservation: A "physics-fidelity" check to see if the models accidentally "create" or "destroy" material during the simulation.

Datasets

The primary focus is the Perturbed Layer Interface (PLI) dataset.

  • Composition: 8 channels (6 material densities + 2 velocity fields).
  • Resolution: 1120 × 400 Eulerian grid points.
  • Scale: 3850 training trajectories and 100 test trajectories, each spanning 101 timesteps.
  • Materials: As shown in Figure 1, the simulation includes a copper throw, aluminum, polymer, high explosive, a steel case, and air.
Initial state of the PLI simulation showing the different material layers.

Key Insights and Ideas

  • Diffusion Models are Surprisingly Robust: Despite not being specifically designed for PDEs, the EDM-L diffusion model was the most performant. It captured the complex deformation of materials better than models specifically built for physics.
  • The "Resolution Trap" for Transformers: Foundation models like POSEIDON struggled significantly. To make the high-resolution data tractable for the transformer, the authors had to increase "patch sizes." This caused the model to lose the fine details of thin material layers, which then disappeared entirely in the predictions.
  • Sparsity Masks Failure: In multi-material problems, most of the grid is "empty" for any given material. A model can achieve a very low error (RMSE) just by predicting zero everywhere. This highlights why structural metrics like Pearson correlation are non-negotiable for evaluating physics emulators.
  • Auto-regressive Stability: Most models struggle with "rollout error"—where small mistakes in step 10 compound into total nonsense by step 100. EDM-L showed remarkable stability in these long-horizon predictions compared to the U-Net and Transformer approaches.

Key Results

  • EDM-L vs. Ground Truth: Qualitatively, EDM-L closely matched the complex deformations seen in the ground truth simulations, even at late timesteps ($t=100$), as shown in Figure 2.
Qualitative comparison between ground truth and EDM-L predictions.
  • Quantitative Edge: EDM-L maintained significantly higher Pearson correlation and lower RMSE on the "steel case" and "copper throw" channels compared to EDM-S and SPUS (see Figure 3 and Figure 4).
  • Speed vs. Accuracy: There is a massive trade-off. SPUS is incredibly fast (1.1 seconds per trajectory) but less accurate. EDM-L takes about 32 minutes per trajectory—still faster than the 45 minutes required for a full 224-core CPU simulation, but the margin is much tighter.
Comparison of pixelwise RMSE across models.

Limitations and Future Work

  • Inference Speed: While EDM-L is accurate, its high inference cost (due to the iterative nature of diffusion) limits its use for real-time applications.
  • Foundation Model Flexibility: Pre-trained PDE models currently lack the flexibility to handle high-resolution data without significant re-architecting, which often destroys the benefits of pre-training.
  • Long-Horizon Challenges: Even the best models show increasing error over time. Future work needs to focus on enforcing physical constraints (like energy conservation) more strictly to prevent these drifts.

The authors conclude that while we are making progress, the path to truly general-purpose PDE foundation models requires architectures that can natively handle the high-resolution, multi-material sparsity found in real-world engineering.

Summary updated

Abstract

Machine learning (ML) architectures are being developed for fast emulation of numerical simulators to accelerate both scientific discovery and engineering design. While these approaches are well-tested on popular PDE (partial differential equation) benchmarks, it is necessary to evaluate their performance on datasets that more accurately reflect industrial and engineering applications. We evaluate several computer vision-based architectures trained as surrogate models on a challenging computational physics task involving multiple materials, revealing significant implementation challenges for adapting certain pre-trained models, including transformer-based architectures. For such multi-physics multi-material datasets, our results underscore the importance of learning algorithms designed for sparse and imbalanced data, techniques for adapting spatial encodings and physical consistency metrics, in addition to methods for reducing the cost of generating sufficient training data.
BibTeX
@misc{oyen2026application,
  author = {Diane Oyen and Morgan Jones and Adithan Kathirgamaraju and Abu Bucker Siddik and Alexander Most and David Schodt and Christine Sweeney and Kyle Hickmann and Michael McCann and Jemma Stachelek and Alexius Wadell and Michal Kucer and Mahindra Rautela and Siddharth Mansingh and James Amarel and Gerd J. Kunde and Emily Casleton Taylor and Soumi De and Ayan Biswas and Aimee Hungerford and Nathan Debardeleben and Earl Lawrence},
  title = {{Application of PDE emulators for a multi-material, multi-physics simulation problem}},
  year = {2026},
  eprint = {2610.37972},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.37972},
}
Figure