Physics-informed Kolmogorov–Arnold network for fire source localization and temperature field reconstruction in multi-source tunnel fires

Abstract

Tunnel fires with multiple fire sources pose significant challenges for sources localization and temperature field reconstruction. This study proposes a Multi-source Physics-informed Kolmogorov–Arnold Network (PIKAN) to solve this inverse problem using sparse thermocouple data. The proposed framework embeds the tunnel ceiling heat diffusion equation into a KAN-based neural network, in which radial basis function layers are incorporated to enhance nonlinear approximation capacity and ensure stable automatic differentiation. Specifically, the network comprises a temperature field network and multiple heat release rate networks, which are jointly trained to infer fire source locations, heat release characteristics, and the spatiotemporal temperature rise field. The method is validated using a full-scale double-source tunnel fire experiment and a three-source tunnel fire scenario simulated by Fire Dynamics Simulator. Compared with the conventional Multi-source PINN, PIKAN achieves higher localization accuracy and more reliable temperature reconstruction. In the double-source case, the localization errors are reduced to 0.89 m and 1.06 m, respectively. In the three-source case, the localization errors are controlled within 1.07 m, with the minimum error reaching 0.02 m. The reconstructed temperature fields agree well with experimental and numerical results. The results demonstrate that Multi-source PIKAN provides an accurate and physically consistent approach for multi-source fire inversion in tunnel fires.