An enhanced deep learning method for accurate and robust modelling of soil stress-strain response

The representation of soil stress-strain response by using neural networks has received considerable attention as a promising data-driven method. Recently, a magnitude-related accuracy issue on stress-strain response was exposed for the neural network-based method, where the accuracy had an apparent decay when predicting the low-magnitude stress and strain data. This study proposes an enhanced deep learning method to tackle this issue by the fair reallocation of weight gradient. The enhanced method can significantly improve the accuracy, extrapolation capacity, robustness of neural network-based methods. A rationality investigation is also conducted via an insight into the weight gradient variation in neural networks. The effectiveness of this enhanced method is verified by three stress-strain responses of soil: a raw synthetic stress-strain response for accuracy assessment, a noised synthetic response with Gaussian noise for robustness, and a limited measured response from laboratory test. In predictive performance, the enhanced method improves the predictive accuracy of the stress-strain response with the mean absolute percentage error (MAPE) < 3% for raw and noised responses, MAPE < 6% for laboratory dataset. The improved extrapolation capacity and robustness against errors are also discussed.