Data-driven surrogates for predicting thermal performance and response functions of thermo-active piles
Javier Sánchez Fernández, Agustín Ruiz Lopez, David TabordaPurpose
This paper develops fast, accurate data-driven surrogate models to predict the thermal performance and thermal response functions of single thermo-active piles. These replace computationally expensive finite element simulations and geometry-specific g-function calculations with generalisable machine-learning models suitable for preliminary design and performance assessment. This study aims to improve accessibility, reduce computational cost and support wider adoption of thermo-active piles within low-carbon and net-zero infrastructure strategies.
Design/methodology/approach
Two Artificial Neural Network surrogates were trained on databases generated from 3D transient finite element simulations of thermo-active piles. One surrogate predicts transient power output per unit length under a prescribed inlet fluid temperature, while the second predicts normalised pile wall and outlet thermal responses under constant heat flux. Input parameters were sampled using Latin Hypercube sampling. Model training used feature normalisation, cross-validation and regularisation, with performance evaluated using standard regression metrics and SHAP-based interpretability analysis.
Findings
Both surrogate models demonstrate excellent predictive accuracy and strong generalisation. The power output surrogate achieves R² values exceeding 0.99 with mean absolute errors typically below 2 W/m for most of the operational period. The thermal response surrogate reproduces pile wall and outlet g-functions over 10 years with global per-point R² values of 0.994–0.997 and per-timestep averages of 0.971 (wall) and 0.959 (outlet). Validation against a field thermal response test confirms reliable extrapolation beyond the trained diameter range. Computational time is reduced by several orders of magnitude compared with finite element analysis.
Originality/value
This study presents the first generalisable surrogate framework capable of predicting both power output per unit length and pile-specific g-functions for single thermo-active piles across a wide parameter space. By combining 3D numerical simulations with machine-learning surrogates and interpretability analysis, the work bridges the gap between physics-based modelling and practical engineering design. The approach provides a computationally efficient alternative to traditional thermo-active pile design methods.