Accurately predicting weather-driven dynamics of soil desiccation cracks is crucial for quantifying the degradation of soil mechanical and hydrological properties, but it remains challenging and unresolved due to the complex dynamic features of desiccation cracks with different climate variables. Existing physical methods often adopted moisture-dependent, single-variable and fixed constitutive functions to describe the crack dynamics, always leading to insufficient prediction results. In this study, a novel moisture-independent deep learning model incorporating different climate variables was proposed to predict the dynamic changes of desiccation cracks by constructing multi-input-output long short-term memory neural networks (M- -LSTM). A soil column test under long-term artificial weather conditions was conducted to obtain temporal changes of climate variables and crack geometric parameters for model training and validation. Then, the performance of M-LSTM was compared with existing empirical, theoretical and numerical models via different criteria, including MSE, MAE, RMSE and R². The results demonstrate that the proposed M-LSTM effectively captures the dependency between the dynamic changes of desiccation cracks and climate variables, stably and almost perfectly predicting the variations in crack density and width as the climate variables change. For instance, the overall MAE of M-LSTM dropped to 0.03, which is 32.7%, 28.8% and 18.9% away from the prediction results using empirical, theoretical and numerical models, respectively. Our further discussions on the performance of single-input LSTM (without considering climate variables) and single-output M-LSTM (with only one crack geometric parameters being predicted) reveal that reducing input variables only slightly improve the prediction performance. It is still recommended to use M-LSTM to predict weather-driven dynamics of soil desiccation cracks due to its better generalization ability, robustness, practicality and interpretability.

Desiccation cracks; Weather-driven dynamics; LSTM; Climate variables; Moisture-independent prediction method