Nowadays, the use of new compound chemical stabilizers to treat marine clay has gained significant attention. However, the complex non-linear relationship between the influencing factors and the unconfined compressive strength of chemically treated marine clay is not clear. In order to study the influence of various factors (dosage, type of stabilizer, curing age) on the unconfined compressive strength of solidified soil during chemical treatment, experiments were performed to determine the unconfined compressive strength of soft marine clay modified with various types of stabilizers. Further, an artificial neural network (ANN) model was used to establish a prediction model based on the unconfined compressive strength test data and to verify the performance. Sensitivity and optimization analyses were further conducted to explore the relative significance of parameters as well as the optimal dosage amount. Research has found that when the content of aluminate cement is 89.5% and the content of curing agent is 30%, the unconfined compressive strength significantly increases after 28 days of solidification, and the change in quicklime content has the greatest effect on the improvement in the unconfined compressive strength. The influence of modifiers on the unconfined compressive strength is in the order: potassium hydroxide > kingsilica > quick lime > bassanite. The values of each factor were obtained when the unconfined compressive strength was the maximum, which provided support for the optimization of the treatment scheme. The analysis of chemical treatment is no longer limited to the linear relationship according to the test results, which proves the feasibility of non-linear relationship analysis based on the artificial neural network.

The load transfer method is important for the settlement prediction of axially loaded piles, but in multi-layered complex soils, it lacks analytical solutions. Traditional numerical methods such as the finite element method suffer from strong dependence on mesh generation, time-consuming iterative calculations, and high computational costs for back-analysis. This paper proposes a load transfer analysis model based on a Domain Decomposition Physics-Informed Neural Network. A multi-subnet parallel architecture is adopted to simulate multi-layered soils, solving the problem of inter-layer stress–strain discontinuity through interface coupling and gradient continuity constraints; a non-dimensionalization system and a hard constraint mechanism are introduced to enhance training efficiency and physical consistency; and a two-stage analysis framework comprising surrogate model forward analysis and field data inversion is established. Numerical experimental results indicate that the forward analysis of this model is in high agreement with FEM simulation results, and computational efficiency is improved by six orders of magnitude; based on a small amount of field static load test data, multi-layer soil parameters are accurately inverted, achieving more precise pile settlement prediction than FEM. Comparative analysis validates the effectiveness of the domain decomposition multi-subnet over a single network, demonstrating extensibility to hyperbolic and exponential multi-soil constitutive models.

Pile foundations are commonly used to support structures subjected to complex loading conditions. In seismic-prone regions, understanding the soil–pile interaction under cyclic loading is essential for ensuring the stability and safety of these foundations. Numerical modeling is an effective tool for predicting the nonlinear behavior of soil under seismic excitation, but selecting an appropriate constitutive model remains a significant challenge. This study investigates the seismic behavior of pile groups embedded in soft clay using advanced finite element analysis. The piles are modeled as aluminum with a linear elastic response and are analyzed within a soil domain characterized by two kinematic hardening constitutive models based on the Von Mises failure criterion. Model parameters are calibrated using a combination of experimental and numerical data. The study also examines the influence of pile spacing within the group on seismic response, revealing notable differences in the response patterns. The results show that the nonlinear kinematic hardening model provides a more accurate correlation with experimental centrifuge test results compared to the multilinear model. These findings contribute to enhancing the understanding of soil–pile interaction under seismic loading and improving the design of pile foundations.