Why would a deep learning surrogate trained on fixed detector geometry perform worse under size changes than spacing redistributions?

# Why Fixed Geometry Surrogates Struggle with Size Changes

This is a great question about the practical limits of deep learning surrogates in physics—and the papers here offer concrete insight.

The core issue comes down to what the model actually learns versus what changes when you alter detector geometry.

The fundamental asymmetry

When you change *detector size* (e.g., make calorimeter layers thicker or expand the active volume), you're altering a global property that affects how particles propagate through the entire system. A particle shower develops differently in a larger detector: photons travel farther before being absorbed, electrons undergo more scattering, energy spreads over different spatial scales. The model trained on the original geometry has implicitly learned those specific patterns of energy deposition and shower development [AG-2025.11-1571].

By contrast, when you *redistribute detector element spacing* (e.g., change the granularity of readout channels while keeping the same total volume), you're modifying a local, sampling-level detail. The underlying physics of particle interactions stays the same; only the resolution at which you record it changes. Think of it like changing the zoom level of a camera—the scene is identical, just captured at different magnification.

Why spacing is more resilient

A deep learning model trained on fixed geometry learns representations that are somewhat invariant to sampling density because the physical processes (ionization, bremsstrahlung, multiple scattering) operate at fixed scales independent of detector granularity [AG-2025.02-1472]. Spacing redistributions are essentially interpolation or downsampling problems—tasks neural networks handle well. The point cloud representation used in modern surrogate approaches naturally accommodates variable spacing, since it encodes particle positions directly rather than assuming a fixed voxel grid [AG-2025.11-1571].

Size changes, by contrast, require the model to extrapolate the development of particle showers into a regime it never saw during training. The shower's mean free path, diffusion length, and energy branching tree all scale with detector dimensions. A model has no principled way to generalize this without retraining.

Evidence from transfer learning

The papers demonstrate this gap clearly: transfer learning with 100 new samples and spacing-like modifications achieves 44% improvement over retraining from scratch, while full geometry transfers still require substantial fine-tuning [AG-2025.11-1571]. Similarly, switching between similar collider detectors (CLICdet to CLD) works well because the underlying geometries are comparable; the model is adapting readout patterns, not fundamental shower physics [AG-2025.02-1472].

Why this matters: For future detector design, it suggests you can safely optimize element placement and granularity with trained surrogates, but any conceptual redesign (larger detector, different material, new acceptance region) demands either retraining or a more geometry-agnostic architecture like those using mutual information objectives [AG-2025.03-1551].