Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence

Researchers successfully deployed a physics foundation model trained on simulations to predict laboratory turbulence behavior, achieving zero-shot generalization to experimental data without direct exposure to lab conditions. The model resolved a decades-old discrepancy between simulated and experimental Rayleigh-Taylor instability measurements, suggesting initial conditions—not fundamental physics—explain the sim-experiment gap.

This research represents a significant validation milestone for foundation models in scientific domains where simulation-to-reality transfer has historically been problematic. The Rayleigh-Taylor instability presents an ideal test case: it exhibits a well-documented three-fold discrepancy between laboratory measurements and simulations that has remained unresolved despite over a century of investigation. By training the Walrus foundation model on minimal DNS data and applying it zero-shot to messy laboratory conditions, researchers demonstrated emergent capabilities that transcend training distribution—the finetuned model spontaneously entered the experimentally observed regime without ever encountering real experimental samples.

This success challenges conventional machine learning assumptions about domain transfer. Traditional approaches require extensive labeled data from target domains, yet foundation models appear to capture generalizable physical principles that translate across conditions. The zero-shot performance on stable stratification—a regime entirely absent from training—further indicates the model learned compositional physics rather than memorizing specific scenarios.

For the scientific computing community, these results validate foundation models as tools for bridging persistent simulation-experiment gaps. The data-driven evidence that initial conditions drive the alpha discrepancy provides new investigative angles for fluid dynamics researchers. However, the work involves academic physics research with limited direct commercial implications for cryptocurrency or fintech markets. The broader relevance lies in demonstrating foundation models' reliability for critical scientific applications where stakes are high and simulation-reality gaps prove costly.

Future work should expand testing across additional fluid instabilities and experimental apparatus to determine generalization limits and identify failure modes for real-world deployment.

• →Foundation models trained on physics simulations can generalize to experimental data without direct exposure to lab conditions.
• →The Rayleigh-Taylor instability sim-experiment gap of 3x appears driven by initial conditions rather than fundamental physics discrepancies.
• →Zero-shot performance on unseen physical regimes demonstrates foundation models capture compositional physical principles beyond memorization.
• →Minimal finetuning (three DNS realizations) proved sufficient to transfer models from idealized simulations to sparse, noisy laboratory data.
• →This validates foundation models as tools for resolving longstanding discrepancies in scientific experiments and accelerating physics discovery.