Physics-informed Deep Learning to Solve Three-dimensional Terzaghi Consolidation Equation: Forward and Inverse Problems

He Wang (王鹤), Biao Yuan, Ana Heitor, He Wang, Xiaohui Chen

Abstract

This study introduces a Physics-Informed Neural Network (PINN) framework to address the three-dimensional Terzaghi consolidation equation, a cornerstone of geotechnical engineering. By integrating physical laws directly into the neural network's loss function, the proposed method provides a mesh-free solution for both forward prediction of soil settlement and inverse identification of soil properties. Results demonstrate over 99% accuracy and a significant reduction in inference time compared to traditional numerical methods, even in the presence of substantial data noise.

Introduction

In geotechnical engineering, consolidation—the process by which soil settles as excess pore water pressure dissipates—is critical for the stability of infrastructure. While one-dimensional consolidation is well-documented, real-world scenarios involve complex three-dimensional (3D) multidirectional flow. Traditional numerical methods like the Finite Element Method (FEM) and Finite Difference Method (FDM) are computationally expensive, sensitive to mesh quality, and struggle with inverse problems, such as determining the coefficient of consolidation ($C_v$) from field data.

This paper proposes a PINN-based approach to overcome these limitations. PINNs leverage automatic differentiation to satisfy governing partial differential equations (PDEs) without requiring a predefined mesh, offering a flexible and efficient alternative for complex 3D geotechnical problems.

Problem Formulation

The study models a 3D block of consolidating soil subjected to a uniform load $p_0$. The domain is defined by dimensions $l, b,$ and $h$, with specific boundary conditions (top, bottom, left, right, front, and back) that dictate drainage and pressure.

3D domain and boundary conditions for the consolidating soil model.

Methodology

The core of the methodology is a deep neural network that takes spatial coordinates ($x, y, z$) and time ($t$) as inputs to predict the excess pore water pressure ($u$).

The Physics-Informed Neural Network (PINN) architecture for solving the 3D Terzaghi equation.

The architecture consists of four hidden layers with 40 neurons each, utilizing the Tanh activation function. The total loss function is a weighted sum of three components:

  1. Initial and Boundary Condition Loss ($L_c$): Ensures the model respects the physical constraints at the domain edges and start time.
  2. Governing Equation Loss ($L_f$): Penalizes deviations from the 3D Terzaghi consolidation equation using automatic differentiation.
  3. Data Loss ($L_d$): Minimizes the error between predictions and available sensor or reference data.

Optimization is performed using a hybrid approach: the Adam optimizer for initial global search followed by L-BFGS for fine-tuning.

Key Results

The PINN framework was evaluated on both forward and inverse problems using synthetic data generated via FDM.

  • Forward Problem Accuracy: The model achieved a relative $L_2$ error of approximately $10^{-3}$, indicating over 99% accuracy in predicting the pressure field compared to numerical benchmarks.
  • Inverse Problem Robustness: The PINN successfully identified the coefficient of consolidation ($C_v$) with 99.3% to 99.9% accuracy under low-noise conditions. Remarkably, even with 30% Gaussian noise added to the training data, the model maintained over 95% accuracy in parameter discovery.
  • Computational Efficiency: While training requires approximately 30 minutes, the inference phase is nearly instantaneous. The PINN predicts the entire 3D pressure field in 0.52 seconds, roughly 100 times faster than the 54 seconds required by the numerical solver.

Discussion and Conclusion

The application of PINNs to 3D Terzaghi consolidation represents a significant advancement in geotechnical modeling. By treating physical properties as learnable parameters, the framework simplifies the inverse problem, which is traditionally a tedious trial-and-error process.

However, challenges remain. The "curse of dimensionality" still impacts computational requirements for very large-scale problems, and balancing the various loss terms requires careful tuning. Future work should focus on validating this framework with sparse, heterogeneous data from actual construction sites to bridge the gap between theoretical models and real-world engineering applications.

Summary updated

BibTeX
@inproceedings{2025physics,
  author = {He WANG (王鹤) and Biao Yuan and Ana Heitor and He Wang and Xiaohui Chen},
  title = {{Physics-informed Deep Learning to Solve Three-dimensional Terzaghi Consolidation Equation: Forward and Inverse Problems}},
  booktitle = {Arabian},
  year = {2025},
  doi = {10.1007/s13369-025-10602-2},
}
Figure