Physics-Informed Neural Networks for Radial Consolidation of Combined Electroosmotic, Vacuum and Surcharge Preloading Considering Smear Effects

Researchers develop physics-informed neural networks (PINNs) to model electroosmotic soil consolidation with combined loading conditions. The study compares three neural network architectures, finding that hard-constraint boundary encoding significantly improves accuracy for complex time-dependent loading scenarios, achieving prediction errors under 0.5 kPa.

This research advances computational methods for geotechnical engineering by applying machine learning to complex soil mechanics problems. Physics-informed neural networks represent a bridge between traditional finite element analysis and modern deep learning, enabling faster predictions while maintaining physical consistency. The study addresses a practical engineering challenge—predicting soil consolidation under electroosmotic, vacuum, and surcharge loading—where analytical solutions are intractable and numerical simulations are computationally expensive.

The comparison of three PINN architectures reveals important lessons about neural network design for physics-based problems. Standard soft-constrained approaches struggle with competing objectives in time-dependent scenarios, while the modified hard-constraint variant embedding boundary conditions directly into the network output substantially improves performance. This architectural innovation has broader implications for physics-informed machine learning across engineering disciplines.

The framework's robustness across practical parameter ranges suggests real-world applicability. Achieving sub-killopascal accuracy in pressure predictions could enable faster design iterations for ground improvement projects, reducing engineering consulting time and costs. The dimensional analysis foundation makes the models potentially transferable across different soil conditions and loading magnitudes.

This work exemplifies how AI accelerates specialized engineering domains beyond consumer-facing applications. As computational geotechnics increasingly leverages machine learning, practitioners may shift from expensive FEM simulations toward validated neural network surrogates for preliminary design phases. The sensitivity analyses demonstrating robustness strengthen confidence in deployment potential.

• →Hard-constraint boundary encoding reduces optimization complexity and improves PINN accuracy for soil consolidation modeling
• →Gated neural network architectures better resolve steep pressure gradients near material interfaces
• →The framework achieves sub-0.5 kPa mean absolute error under multiple time-dependent loading conditions
• →Physics-informed neural networks can serve as computationally efficient surrogates for expensive finite element simulations
• →Network robustness across practical parameter ranges supports real-world geotechnical engineering applications