The increasing integration of artificial intelligence (AI) in engineering disciplines is redefining the way we formulate, solve, and interpret complex physical problems. Among these innovations, Physics-Informed Neural Networks (PINNs) have emerged as a powerful tool capable of embedding physical laws, as sets of partial differential equations (PDEs), into deep learning architectures. PINNs have demonstrated remarkable results in solving forward and inverse problems across fluid mechanics, heat transfer, and material modelling, even with sparse or noisy data. However, their application in geomechanics remains in its early development, especially for complex, non-linear, and time-dependent processes such as soil consolidation. Soil consolidation, a fundamental problem in geotechnical engineering, describes the time-dependent settlement associated to the dissipation of excess pore water pressure, induced, for instance, by mechanical loading or change in hydraulic boundary conditions. While classical Terzaghi 1D theory provides a simplified analytical framework, practical applications often involve complex boundary conditions, nonlinear material behavior (e.g., viscosity, plasticity and void ratio-dependent properties), and multi-dimensional effects. These aspects pose significant challenges to traditional formulations, as well as to computational methods. The goal of this work is to critically assess the performance of PINNs in modelling consolidation problems and to compare their capabilities and limitations with those of traditional numerical approaches. By using the consolidation problem as a test case, we aim to evaluate how PINNs handle typical issues of geotechnical applications, like discontinuous initial/boundary conditions, and to provide methodological insights for their effective use in soil mechanics. This study focuses on identifying some key factors that may influence the reliability of the PINN results, such as loss function design, hyperparameter sensitivity, and the evaluation of the evolution in space and time of the stress and the displacement fields as derived from the output of the PINN, i.e. the excess pore water pressure.
Challenges, merits and limits of PINN applied to clay consolidation processes
P. Marveggio;M. Cremonesi;G. Della Vecchia;
2025-01-01
Abstract
The increasing integration of artificial intelligence (AI) in engineering disciplines is redefining the way we formulate, solve, and interpret complex physical problems. Among these innovations, Physics-Informed Neural Networks (PINNs) have emerged as a powerful tool capable of embedding physical laws, as sets of partial differential equations (PDEs), into deep learning architectures. PINNs have demonstrated remarkable results in solving forward and inverse problems across fluid mechanics, heat transfer, and material modelling, even with sparse or noisy data. However, their application in geomechanics remains in its early development, especially for complex, non-linear, and time-dependent processes such as soil consolidation. Soil consolidation, a fundamental problem in geotechnical engineering, describes the time-dependent settlement associated to the dissipation of excess pore water pressure, induced, for instance, by mechanical loading or change in hydraulic boundary conditions. While classical Terzaghi 1D theory provides a simplified analytical framework, practical applications often involve complex boundary conditions, nonlinear material behavior (e.g., viscosity, plasticity and void ratio-dependent properties), and multi-dimensional effects. These aspects pose significant challenges to traditional formulations, as well as to computational methods. The goal of this work is to critically assess the performance of PINNs in modelling consolidation problems and to compare their capabilities and limitations with those of traditional numerical approaches. By using the consolidation problem as a test case, we aim to evaluate how PINNs handle typical issues of geotechnical applications, like discontinuous initial/boundary conditions, and to provide methodological insights for their effective use in soil mechanics. This study focuses on identifying some key factors that may influence the reliability of the PINN results, such as loss function design, hyperparameter sensitivity, and the evaluation of the evolution in space and time of the stress and the displacement fields as derived from the output of the PINN, i.e. the excess pore water pressure.| File | Dimensione | Formato | |
|---|---|---|---|
|
Marveggio_et_2025_fomlig.pdf
accesso aperto
:
Publisher’s version
Dimensione
381.37 kB
Formato
Adobe PDF
|
381.37 kB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.



