A machine learning assisted probabilistic framework for predicting the impact of crack characteristics on heterogeneous deformation of desiccated soils under environmental changes

• Monte Carlo crack-width distributions closely matched multi-physics simulations. • ML surrogate model reproduced crack widths with low prediction error. • FOSM efficiently estimated crack statistics and ranked influential parameters. • Risk analysis showed 92% subsidence vs 8% swelling under cyclic climate effects. • Increasing initial crack width reduced subsidence probability in risk space. Progressive surface deformation of desiccated cracked soils in semi-arid to arid regions in response to seasonal changes has become a challenge to the performance of infrastructures. Hence, a comprehensive analysis of heterogeneous deformation due to changes in crack width was conducted by considering more reliable cumulative distributions for environmental parameters. The probabilistic modelling was performed and validated using data from field reconnaissance and numerical simulations to establish relationships between crack width, spacing, and surface deformation for different crack geometries. Monte Carlo sampling and the first-order second moment method were employed to determine probability distributions and evaluate subsidence risks. Moreover, a machine learning approach was employed to elucidate the relationships between relevant parameters. As a result, the significance of soil-atmosphere interplay for various parameters from the most to the least important was revealed for cracked ground surfaces. Relative humidity, precipitation, and radiation emerged as the most critical factors affecting deformation of cracked soil. The probabilistic estimation of crack widths and the reliability index are used to assess the likelihood that crack sizes will remain within acceptable limits, even under changing climatic conditions. This also accounts for the associated risks of ground subsidence and swelling caused by variations in crack widths.