Abstract
Comprehensive tunnel lighting assessment requires three-dimensional illuminance distributions across the roadway and surrounding surfaces. Physically based simulation resolves these distributions but becomes costly during repeated evaluations, while data-driven models commonly focus on individual planes. Here we present TunnelLuxNet, a conditional CNN–Transformer framework for rapid prediction of tunnel interior illuminance fields. A geometry-aware encoding maps connected surfaces to a common layout represented by multichannel spatial fields and a condition vector. Within this representation, convolutional processing captures local features, axial attention establishes longitudinal and cross-surface interactions, and conditional modulation incorporates geometric, material and luminaire conditions. Trained on 3,000 Radiance-generated configurations, TunnelLuxNet achieves mean MAE and RMSE of 0.097 and 0.404 lx for local fields and 1.290 and 2.286 lx under the densest tested installation, with close agreement against published measurements. Predicting approximately 1.8 million surface illuminance values for a long tunnel takes 22 s, compared with an estimated 6 h using Radiance. Extrapolation tests reveal greater sensitivity to cross-sectional geometry and luminaire aiming than to reflectance or alignment. Separately trained models for batwing and narrow-beam photometries demonstrate the workflow’s applicability to different luminaires. Ablation results highlight cross-surface information exchange as a key contribution to accuracy, suggesting its value for learning coupled physical fields in three-dimensional environments. TunnelLuxNet provides a reusable computational basis for comprehensive lighting assessment and extensive evaluation of design alternatives.

This publication has 20 references indexed in Scilit: