Abstract
<title>Abstract</title> <p>Spectral-robust training in diffractive neural networks may depend on phase parameterization, propagation sampling, and device conditions. We compared sample-wise worst-case and average multi-shift objectives in the same dual-wavelength differential architecture, using bounded fused-silica thickness and Malitson Sellmeier dispersion. A broad benchmark comprised eight prespecified configurations with ten independent training runs each. The primary comparison further used ten paired runs per objective initialized and optimized directly on a 128 × 128 propagation grid, followed by a prespecified physical-domain analysis. In direct 128-grid training, the paired difference in balanced wavelength-robustness score area under the curve (Balanced-WRS AUC; worst-case minus average-loss) was − 0.00537 (bootstrap 95% CI, − 0.02080 to 0.01143; exact two-sided sign-flip p = 0.5469). The prespecified ± 20 nm training-domain confirmation yielded − 0.00557 (95% CI, − 0.02131 to 0.00833; p = 0.4980). Neither confirmatory comparison detected a reliable worst-case advantage. Point estimates changed sign between frozen-device resampling and direct high-resolution optimization, but all corresponding intervals crossed zero; this suggests possible protocol sensitivity without establishing a general law. Independent per-layer displacement reduced performance toward chance, further limiting the scope of spectral-robustness claims.</p>