Evaluating the safety impacts of three-dimensional alignments of mountainous freeways using an interpretable machine learning approach

The safety of mountainous freeways remains a critical concern in road infrastructure planning, especially for routes characterized by intricate geometric alignments. The roadway is essentially a ribbon-shaped structure in three-dimensional (3D) space, and using 3D spatial parameters allows for a more accurate representation of its true geometric characteristics. However, quantifying the impact of 3D alignment features on crash occurrence remains challenging. To this end, this study investigates the nonlinear relationship between crash rate and 3D alignments using 4 neural network models: multilayer perceptron (MLP), Wide & Deep, residual network (ResNet), and graph neural network (GNN). The crash dataset was collected from two mountainous freeways in Guangdong Province, including crash records, traffic volumes, and road geometry over a 3-year period (2021–2023). According to model evaluation metrics, the Wide & Deep model achieved the best predictive performance. To enhance model interpretability, the SHapley additive exPlanations (SHAP) method and partial dependence plot (PDP) method were applied, enabling insights into the global importance of each feature and their interaction effects. SHAP results indicate that three geometric attributes, namely vertical curve length (VC_length), horizontal curve length (HC_length), and grade, have a significant influence on crash rate. Notably, 3D curvature-related features emerged as key predictors. As shown in the SHAP analysis, both excessively high and low curvature values, as well as a pronounced drop in curvature between consecutive segments, exhibit negative SHAP contributions, implying a mitigating influence on the predicted crash risk and underscoring the significance of curvature consistency. By contrast, torsion-related variables show limited impact. Additionally, key interaction effects were identified among several variable pairs, particularly between VC_length and HC_length, VC_length and back curvature difference, and HC_length and back curvature difference. These findings provide practical guidance for optimizing geometric design on mountainous freeways, supporting safer and more informed infrastructure planning.