8 Dec 2025

Improved Prediction of Soil Thermal Properties Using Gated Recurrent Unit Neural Networks

Abstract: Frost actions, such as frost depth penetration and thaw weakening, are damaging to airfields and roadways in cold regions. Machine learning techniques, such as recurrent neural networks, have been applied to this problem, but with a large focus on long short term memory (LSTM) neurons. Gated recurrent units (GRUs) are similar to LSTM neurons in terms of accuracy, but are more computationally efficient, and have yet to be applied to predicting soil thermal properties. Using a hyperparameter search, an optimal architecture for a recurrent neural network based on gated recurrent units was identified. A general model using temperature, thermal conductivity, and volumetric moisture content was found to predict temperatures effectively, having an error of less than 0.25°F across all depths. For predicting thermal conductivity, a model including temperature but not moisture content was found to be effective. For moisture content, the results were inconclusive as both models were affected by similar errors. Overall, the GRU-base recurrent neural networks were found to work well for predicting soil thermal properties in high-plasticity clays, and it is recommended to further expand the training datasets to include other frost-affected soil types.