2026 Volume 7 Issue 3 Pages 879-889
This study employed a Physics-Informed Neural Network (PINN) to identify empirical physical parameters in a one-dimensional shallow-layer model for dense-gas gravity currents and evaluated its predictive performance through comparisons with laboratory-scale experiments. A Froude-number-based scaling was introduced to establish a unified PINN framework applicable to different initial dam-height conditions. Forward analyses were then performed using the physical parameters identified by inverse analysis. The PINN successfully reproduced the temporal evolution of the gas front position and floor concentration under different initial conditions. In contrast, for the vertical concentration profiles, the PINN tended to underestimate the concentration at unseen locations where large experimental trial-to-trial variations were observed. These results demonstrate the effectiveness and applicability of the proposed PINN framework for predicting the spatiotemporal behavior of dense-gas gravity currents.