Abstract: Physics-informed neural networks (PINNs) have attracted considerable attention in geotechnical engineering as an artificial intelligence approach for inverting soil and rock parameters by integrating measurement data with physical laws. However, due to the highly nonlinear mapping between parameters and the loss function, PINNs often fail in complex, high-dimensional parameter inversion problems. To enhance the generalization capability of PINNs, a damped physics-informed neural network (dPINN) is developed by improving the conventional PINN, and is applied to invert the hydraulic conductivity field of geologic media—spatially following a generalized sub-Gaussian distribution—in two cases: a two-dimensional steady-state saturated groundwater flow (Case 1) and a one-dimensional transient soil infiltration (Case 2). The results show that: (i) the dPINN can effectively invert the conductivity field in both cases, reducing the relative L₂-norm error (Y) in Case 1 by 29% compared with the conventional PINN; (ii) as the non-Gaussianity of the conductivity field decreases, Y in Case 1 decreases from 0.88 to 0.71; and (iii) the sensitivity of inversion accuracy to the weights of different physical constraints ranks as governing equation > Dirichlet boundary > Neumann boundary. The proposed method and findings provide a useful reference for geotechnical numerical modeling and parameter inversion.

Key words: groundwater flow, damping physics-informed neural networks, hydraulic conductivity, parameter inversion