FF-PINN slashes geotechnical simulation errors by 66% in displacement, halves training time
A new neural network approach cuts FEM-like accuracy in half the time for soil mechanics.
Physics-Informed Neural Networks (PINNs) have long struggled with spectral bias, particularly in high-gradient regions like elastic-plastic boundaries in geotechnical problems. To address this, researchers from Thailand—Apisit Robjanghvad and Sompote Youwai—introduce Fourier Feature PINNs (FF-PINN) for the non-associative Mohr-Coulomb model. By embedding random Fourier feature mapping into the input layer, the network learns high-frequency components more effectively, supported by a multi-objective loss function that enforces equilibrium, constitutive laws, and Karush-Kuhn-Tucker conditions against FEM data. A strain-adaptive sampling strategy further sharpens accuracy in localized plastic zones.
Benchmarked across three test cases, FF-PINN achieves error reductions of up to 66% in displacement fields and 27% in stress components compared to standard PINNs, while faithfully reproducing plastic failure zone geometry. Crucially, the method converges in half the training epochs, cutting wall-clock training time dramatically. Sensitivity analysis confirms robustness across training data size, collocation density, loss weighting, and noise levels up to 2.0%. This work provides a practical, physics-consistent alternative to the computationally expensive Finite Element Method for elasto-plastic geotechnical analysis, enabling faster simulations without sacrificing fidelity.
- FF-PINN reduces displacement prediction errors by up to 66% and stress errors by 27% versus conventional PINNs.
- Converges in half the training epochs, halving wall-clock training time while maintaining higher accuracy.
- Uses random Fourier feature mapping to overcome spectral bias, handling sharp gradients at elastic-plastic boundaries.
Why It Matters
FF-PINN offers a faster, accurate mesh-free solver for complex geotechnical simulations, reducing computational costs.