Application of Deep Neural Operators for Thermal Response Test Analysis: A Data-Driven Approach for Ground Heat Exchanger Characterization

Abstract

Thermal response tests are essential for determining ground thermal properties required for optimal design of ground heat exchanger systems in shallow geothermal applications. This paper presents a novel Multiple-Input Deep Neural Operator variant for rapid and accurate thermal response test analysis across standard, constant temperature, and oscillatory protocols. The approach directly maps heat injection and fluid temperature time series to soil thermal conductivity, borehole thermal resistance, and (for the oscillatory case) soil volumetric heat capacity, without requiring iterative optimization. Models trained on synthetic samples generated by simulations generalize to six real-world datasets from Canada, Denmark, Japan, and the United States. For two-parameter inversion, mean absolute percentage errors are 4.98% for soil thermal conductivity and 2.75% for borehole thermal resistance. In the three-parameter oscillatory case, errors are 0.58% and 5.79% for the same parameters and 9.33% for soil volumetric heat capacity. The models cut interpretation time to seconds and generalize across different input functions without retraining, making it particularly valuable for rapid site assessment and ground-source heat pump system optimization. Future extensions include noise augmentation and physics-informed training.