Numerical Recovery of Pore-Air Pressure and Effective-Stress Reduction in Cover Soils Using a Two-Phase Inverse Physics-Informed Neural Network

Transient pore-air pressure in gas-loaded cover soils is difficult to observe between monitoring depths. We develop a coupled water–air inverse physics-informed neural network as a numerical proof of concept. Under a pre-calibrated constitutive model, three numerically sampled depths supply pore-air pressure, pore-water pressure, and saturation targets. An independent air-pressure field represents overpressure; the coupled balances constrain the joint state. On a one-dimensional same-equation benchmark, the full-window gas-pressure error is 0.0463. The data-only ablation reaches 0.0043, while the coupled residuals improve water pressure, saturation, and front monotonicity. Before the 495 s reference injection-base zero-stress crossing, gas-pressure error is 0.0448; on cells with positive reference effective stress, the stress-reduction error is 0.0450. Later states test numerical tracking under overload. The plane-model case fits two-dimensional FLAC2D training series with a one-dimensional residual omitting lateral transport, giving a responding-depth excess-overpressure error of 0.186. A residual-free line-M test evaluates architecture-only interpolation at 11 held-out depths. The configuration-specific Bishop post-process yields zero-to-seven-minute onset times and a 43% minimum-stress spread across three effective-stress parameter forms. These numerical fields support subsequent mechanical interpretation; physical validation requires measured interior states, and stability assessment requires a mechanical model.