Neural aberration fields: continuous implicit representation of spatially varying point spread functions for vision simulation

Convolution-based human vision simulation requires point spread functions (PSFs) that vary continuously across a high-dimensional parameter space encompassing field angle, pupil diameter, and eye-specific anatomy. Current state-of-the-art methods precompute PSFs on a discrete grid and rely on trilinear interpolation at runtime, incurring substantial memory overhead and resolution limitations. We propose Neural Aberration Fields (NAF), a framework that replaces discrete lookup tables with continuous neural implicit representations mapping optical coordinates directly to Zernike aberration coefficients. We systematically evaluate three neural encoding strategies—positional encoding, multiresolution hash encoding, and direct input—conditioned on patient-specific eye parameters, against the conventional trilinear scheme. Training data are generated from a physically grounded pipeline comprising a parametric four-surface eye model and Snell’s-law ray tracing through 200,000 configurations spanning five ocular conditions. The positional-encoding variant (NAF-PE) achieves a Zernike coefficient MAE of 0.0032 mm and PSF PSNR of 53.3 dB while requiring only 139 KB of storage—an 84% memory reduction and 2.1× accuracy improvement over trilinear interpolation (MAE 0.0067 mm, 864 KB). On unseen cataract parameters, NAF-PE reduces MAE by 43% relative to the discrete baseline (0.0051 versus 0.0089 mm), demonstrating the generalization advantage of conditioned implicit representations. A conditioning ablation reveals that the accuracy gains stem primarily from patient-specific conditioning rather than representational capacity alone, identifying conditioning as a structural necessity for continuous PSF modeling. We further show that hash encoding achieves the highest single-configuration PSNR (56.1 dB) when the table capacity is matched to the data density, but degrades sharply when oversized. These results establish neural implicit fields as a viable, compact, and generalizable alternative to discrete PSF grids for real-time vision simulation individualized within parametric eye models.