Transfer Learning Enhanced Composite Physics-Informed Neural Network: Application to Offshore Geotechnical Experiments

Abstract The uniaxial compression test is a critical method widely used to evaluate the mechanical properties and deformation behavior of seabed soils. It plays an indispensable role in understanding soil strength characteristics, compression indices, and failure mechanisms, which are essential for the design and stability analysis of offshore structures. When conducting extensive experiments on seabed soils from various regions, numerical simulation has emerged as an efficient alternative. Recent advancements in deep learning have demonstrated the significant potential of Physics-Informed Neural Networks (PINNs) for simulating this process. In offshore geotechnics, the governing equations, particularly those centered on the solid momentum equation and constitutive equation, are highly complex. The Composite Physics-Informed Neural Network (CPINN) addresses the pathological gradient issues inherent in traditional PINNs architectures. This study proposes a Transfer Learning Enhanced Composite Physics-Informed Neural Network (TL-CPINN) approach. This method selectively applies transfer learning to the general solution network within a pre-trained CPINN, while retaining the original distance and particular solution networks. Without the need for additional datasets, this approach effectively inherits the boundary information and physical knowledge learned by the prior CPINN. Using the numerical simulation of the uniaxial compression test of elastic soils as an example, TL-CPINN achieves rapid simulation of mechanical behavior for different soil materials after a single comprehensive CPINN training session. This research underscores the superior memory capabilities of PINNs compared to other numerical simulation methods, thereby significantly enhancing the efficiency of numerical simulations.