A machine learning-based thermal-mechanical parameter inversion of energy pile considering thermo-mechanical behaviors

The thermal-mechanical properties of rock-soil are indispensable for the geotechnical investigation of energy piles according to design codes. The thermal parameters of the stratum and pile are commonly determined through thermal response tests (TRT), and the mechanical parameters are obtained via sampling or in-situ testing, with some parameters also assigned based on engineering experience. However, on the one hand, TRT and in-situ testing are costly and labor-intensive processes that last nearly one week. On the other hand, the accuracy of personal experience on parameter determination is not guaranteed. Parameter inversion adjusts numerically modelled values (originally from tests/experience) to compensate for sampling effects, stratigraphic variations, and human biases that create simulation errors. The machine learning-based surrogate model, experiencing a surge in popularity in recent years, is a promising solution for inversion acceleration. By training on precomputed numerical simulation results, the surrogate model emulates the system’s behavior at significantly faster speeds, simultaneously reducing computational time and financial costs compared to traditional numerical modelling. This study proposes a novel method for parameter inversion with machine-learning-based surrogate models. A numerical model is first established by replicating a reduced-scale physical test capturing the thermal-mechanical behaviors of the energy pile. The Latin hypercube sampling is subsequently utilized to generate sufficient data for XGBoost model development, where the sensitivity analysis is subsequently carried out for parameter screening. The parameter inversion is then implemented with non-dominated sorting genetic algorithms. The effectiveness of the method is validated using 4 real-case test conditions. The results demonstrate that the surrogate model attains high accuracy with an R² of temperature and axial forces above 0.97 and 0.9, respectively. Lastly, the results of parameter inversion indicate a promising and optimistic prospect for the proposed inversion method, which can reveal the comprehensive parameters of the material to a certain extent. This study presents an efficient method for parameter inversion of energy piles while providing a perspective of fast determination of rock-soil thermal properties.