What happens if you deploy a surrogate beyond its training data's energy range?

# Deploying Surrogates Beyond Training Data: Risks and Validation

Your question touches on a real and urgent problem in modern computational physics: what happens when a learned model—a surrogate—is asked to extrapolate into regimes it never saw during training?

The Core Problem

Surrogate models are neural networks or other machine-learning systems trained to mimic expensive simulations or real experiments. They're fast, which is their value. But they're only reliable where the training data is dense. Push them beyond that domain, and they tend to fail in subtle, confidence-inspiring ways: they don't know what they don't know.

In particle physics, this is especially risky. When searching for rare decays or new physics, you might train a surrogate on simulated jets within a certain energy range—say, 20–100 GeV—then deploy it on real detector data spanning 10–500 GeV. The edges are terra incognita.

How to Know When You've Left Safety

The ATLAS collaboration has developed a direct answer. When they released surrogate models for displaced-particle searches, they had to solve the problem of "unknown unknowns" in reinterpretation [AG-2025.02-1260]. Their surrogate models were paired with a critical feature: Bayesian uncertainties built into the model itself [AG-2024.02-1165]. These uncertainty estimates grow large when the model encounters inputs far from its training distribution, signaling unreliability. This is essential for public reuse—the model essentially says, "I don't know my own answer here."

The Fundamental Limits Question

A deeper worry emerges at the frontier: are your surrogates even capturing the physics correctly *within* their training range? The SURF method addresses exactly this [AG-2025.11-1399]. It validates a generative surrogate by comparing its learned likelihood to a reference model trained on real data, using exact statistical tests. The finding is sobering: some modern generative models (like autoregressive GPTs applied to jets) unphysically exaggerate separation power *within* the training domain, implying they're giving a false picture of what's even achievable.

Practical Mitigation

For high-stakes applications, the best practice is a feedback loop: train the surrogate, use it to propose designs or analyses, then verify predictions against the slow simulator or real experiment [AG-2025.11-530]. This catches pathological extrapolation before it pollutes results. In gravitational wave detector design, researchers loop between surrogate prediction, inverse design, and slow-simulator validation continuously.

---

None of these preprints directly address how to recover or correct predictions made far outside the training range. The archive emphasizes validation and uncertainty quantification as prevention, not remediation.