Title: Bridging the Gap: A Physics Foundation Model Transfers from Simulation to Lab-Based Turbulence

Abstract

The practical utility of physics foundation models in real-world laboratory settings remains a significant open question within the field of scientific machine learning (ML). To address this, we evaluate their capabilities using the Rayleigh-Taylor instability (RTI)—a pervasive and complex fluid instability observed across scales ranging from small-scale laboratory flows to massive supernova explosions. In RTI scenarios, minor perturbations at a density boundary expand into chaotic, multiscale mixing as a lighter fluid accelerates into a denser one.

Conventional ML approaches have struggled with RTI. Despite more than a hundred years of theoretical, computational, and experimental investigation, a persistent discrepancy exists between simulations and physical experiments. Specifically, the late-time mixing growth rate ($\alpha$) observed in most laboratory experiments is approximately 0.06–0.07, which is roughly triple the value obtained from idealized direct numerical simulations (DNS) of ~0.02. The source of this divergence is still a subject of debate. Consequently, RTI serves as a rigorous benchmark for a broader inquiry: Can foundation models trained exclusively on simulation data effectively generalize to the sparse, noisy, and unstructured nature of laboratory environments?

We fine-tuned Walrus, a foundation model designed for continuum dynamics, using three or fewer DNS realizations. This process allowed the model to recover essential RTI physics over extended time series. When applied zero-shot to laboratory data featuring sliding barriers, the fine-tuned model successfully moved beyond the DNS-like regime, aligning with the experimentally observed growth band, despite having never encountered any experimental samples during training. These findings offer independent, data-driven confirmation that initial conditions are pivotal in explaining the long-standing sim-experiment gap in $\alpha$.

Furthermore, the model demonstrated zero-shot generalization to stable stratification—a buoyancy regime not included in the training data—correctly predicting a reduction in mixing-layer growth. Collectively, our results indicate that foundation models can generalize effectively beyond their training datasets, accurately predicting laboratory behaviors and unseen physical regimes. This success opens novel avenues for investigating and resolving enduring discrepancies between simulations and experiments.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC