How do physics-informed constraints like Landau straggling prevent ML models from learning unphysical patterns?

# Physics-Informed Constraints and Unphysical Learning

Physics-informed constraints like Landau straggling act as guardrails that anchor machine learning models to reality—they encode what *must* be true based on first-principles physics, preventing the model from fitting noise or inventing patterns that violate conservation laws and symmetries.

How PHIN-GAN Uses Straggling Functions

The clearest example comes from particle physics simulations. When particles pass through matter, their energy loss follows a specific statistical distribution called the Landau straggling function. Researchers developed PHIN-GAN, a generative model that learns to simulate these interactions [AG-2026.04-1383]. Rather than letting the neural network learn energy-loss patterns freely from data, they analytically derived the exact probability distributions that govern straggling and built these into the model's learning objective. This means the generator must output distributions matching the physics—it cannot learn spurious correlations or unphysical tails in the energy spectrum, even if such patterns were present in the training data by accident [AG-2026.04-1383].

Broader Architecture: Embedding Physics Priors

This principle generalizes. In inverse problems across high-energy physics—from inferring material properties in fusion implosions to extracting QCD parameters from collision data—researchers embed physics constraints at multiple levels [AG-2024.12-1755], [AG-2025.01-1151]:

• Symmetries built into network architecture (e.g., equivariant layers that respect rotational invariance) [AG-2024.10-1172]
• Continuity and conservation laws imposed as loss-function penalties
• Sparsity priors derived from hydrodynamic theory rather than data-driven sparsity [AG-2024.12-1755]

These constraints reduce the model's degrees of freedom: instead of freely fitting 1 million weights, it must fit them *while* satisfying 100 known physical laws. Unphysical solutions automatically incur high loss and are discarded during training.

Why This Matters

Standard deep networks are universal approximators—they can fit anything, including nonsense. A model trained only on likelihood or accuracy may learn that energy sometimes increases in a collision, or that particles tunnel through barriers with wrong probabilities, because the training data is finite and noisy. Physics-informed constraints eliminate whole classes of wrong answers before they can be learned [AG-2025.01-1151]. The result is a model that generalizes better to new regimes (untested energies, parameter ranges) because it's constrained to obey rules that hold everywhere, not just in the training set.

The provided preprints focus on embedding physics in generative models and likelihood estimation; they don't discuss Landau straggling in the broader ML context, but AG-2026.04-1383 is a direct application of your question.