Improved Prediction of Soil Thermal Properties Using Recurrent Neural Networks

Frost effects, such as thaw weakening and frost heave, can significantly degrade pavements. Effective mitigation relies on accurate prediction of soil thermal properties. Current prediction methods use empirical equations or finite element analysis, with recent progress using machine learning. Recurrent neural networks, a type of machine learning model designed for prediction of sequential data, may be appropriate. Further, the model’s output can be fed back into itself recursively to make long-term predictions. Using this method and training data from Hill Air Force Base, Utah, and Air Force Academy, Colorado, we were able to forecast soil parameters including temperature, thermal conductivity, and moisture content for frost susceptible soils. The model had a deviation from experimental values of no more than 10%, with the most significant contributions to accuracy being the use of gated recurrent unit neurons and the incorporation of multiple soil parameters.