A machine learning based surrogate model for predicting the influence of environmental conditions on desiccation crack width

Climate change is of significant interest in various fields due to its far-reaching impacts. One of the notable consequences of environmental change is the increase in the frequency and severity of soil desiccation cracking, which can lead to a range of serviceability limit state problems. Therefore, the influence of environmental factors on variations in the crack width is explored in this study through machine learning techniques. The research focuses on five primary environmental parameters including temperature, wind speed, radiation, relative humidity, and precipitation. These variables were meticulously recorded over a year, from 2015 to 2016, by local meteorological stations to capture the temporal variations. Various regression methods were explored, including Linear, Bayesian, Polynomial, Dummy, and Gaussian regressions. The dataset, comprising environmental parameters and corresponding crack widths, was randomized to prevent bias. A split of 70% for training and 30% for testing was adopted to validate the model performance. The Mean Absolute Error (MAE) is employed as the metric to evaluate the accuracy of the regression models. Subsequently, a numerical thermo-hydro-mechanical modeling approach was utilized to compute the variability of crack width. The results obtained from the machine learning methods were then compared against the findings from a comprehensive multi-physical modeling. According to the results, linear regression emerged as the most suitable model among other regression techniques, demonstrating the lowest MAE. The findings also reveal that the predictive model demonstrates an excellent match with the numerical simulation results. This finding implies the high dependency of the variations in crack width on the environmental parameters considered. Lastly, a new formulation was derived and proposed that can be used to estimate the crack width under specific environmental conditions.