<p>Pupil diameter, as a sensitive indicator integrating photometric stimulation and mental workload, provides an effective bridge between tunnel visual conditions and drivers’ psychophysiological states. This study conducted real-vehicle field experiments in four tunnels with different geometric characteristics. Multiple environmental variables were collected, including visibility, road surface and sidewall illuminance, lighting uniformity, correlated color temperature, tunnel distance, curvature radius, and external time, together with drivers’ pupil diameter measured by an eye-tracking system. The grey wolf optimizer-eXtreme gradient boosting model (GWO-XGBoost) was developed to predict pupil diameter, and its performance was evaluated using MSE, RMSE, MAE, and R² with 5-fold cross-validation. The results show that XGBoost model performed better than a number of baseline models, and GWO based hyperparameter optimization enhances predictive performance compared with grid search and random search, achieving an R² of 0.914 and MAPE of 7.15% on the test set. In order to facilitate the interpretability, the SHapley Additive exPlanations (SHAP) analysis was conducted to quantitatively demonstrate the contributions of features, uncover nonlinear effects and investigate interaction schemes. Visibility was found to be the major influence factor on pupil response, and then color temperature, road surface illuminance, finally geometric parameters. To reconcile data-driven observations with psychophysiological mechanisms, a visibility-driven compensation (VDC) model was suggested by combining the photometric pathway and cognitive pathway. The VDC model described the nonlinear relationship between visibility and pupil diameter, with a high level of fit, which is understandable from the standpoint of theory for tunnel lighting optimization and safety-oriented design.</p>