Could inheriting Geant4's approximations limit a surrogate's physics accuracy?

This is a thoughtful question about a real tension in ML-based physics simulation. Let me work through it with the archive.

The Core Issue

Yes, inheriting Geant4's approximations can limit surrogate accuracy — but it depends on *which* approximations and *what* you're using the surrogate for.

Geant4 itself is not physics ground truth; it's a carefully-tuned simulation of detector response based on decades of experimental calibration and theoretical models. When you train a machine learning surrogate on Geant4 outputs, you're learning Geant4's *biases*, not nature's laws. This matters because Geant4 makes simplifying choices — in hadron interaction models, multiple scattering treatments, energy straggling — that may not match reality in extreme or novel regimes [AG-2026.04-1383].

Where This Cuts Deepest

The most direct evidence comes from validation work. The SURF method [AG-2025.11-1399] revealed that some generative surrogates can actually exaggerate the separation power between physics processes — implying they've learned and amplified Geant4's particular parameterizations rather than capturing the underlying data distribution. This suggests the surrogate hasn't escaped its training ground.

However, there's a critical flip side: surrogates trained on Geant4 *can* still be more accurate for practical purposes if they learn to correct systematic biases. PHIN-GAN [AG-2026.04-1383] demonstrates this by enforcing physics-informed constraints (analytical straggling functions grounded in first principles) during training, allowing it to match Geant4's fidelity at a fraction of the cost — implying the surrogate captured what matters without blindly inheriting all approximations.

Geometry Transfer as a Clue

The cross-geometry transfer learning work [AG-2025.11-1571] is revealing here. The fact that surrogates can adapt to *new detector geometries* with only 100 samples suggests they've learned some generalized physics rather than just memorizing Geant4's response. If surrogates were completely locked into Geant4's approximations, they wouldn't transfer so readily.

The Bottom Line

The risk is real but *not inevitable*. A surrogate blindly trained on Geant4 will inherit its approximations. But surrogates designed with physics-informed constraints — or validated against independent data [AG-2025.11-1399] — can escape that trap and actually improve on Geant4 by learning the underlying patterns. The key is validation: if your surrogate only ever sees Geant4, you have no way to know whether it's learned physics or just memorized approximations.

The archive doesn't directly compare surrogates trained on Geant4 to real experimental data at the same scale, so I can't quantify how much accuracy is lost in practice.