GSPINN: A Graph Sequential Physics-Informed Surrogate for Trip Travel Time Prediction

This document outlines the replication materials for GSPINN, a graph-based surrogate model that emulates the SUMO microscopic traffic simulator to predict trip travel times. It provides the exact code, data, and road-network inputs needed to reproduce all tables and figures. The study evaluates model robustness across four distinct SUMO benchmark scenarios (Ingolstadt, Cologne, MoST, LuST) varying in scale, topology, and traffic dynamics. Datasets were generated by simulating traffic states under joint demand-scaling and signal-timing perturbations, capturing diverse congestion regimes using only field-observable attributes. Metrics from a physics-loss weight (λ) sweep are included to regenerate diagnostics. Each network folder contains a full pipeline: data generation, baseline models (SVR, Ridge, MLP, GNNs, Transformers), ablation studies on route pooling, SHAP interpretability plotting, and structural Monotonic GNN baselines (for Ingolstadt and LuST).