AbstractThe swelling of clay-sulfate rocks is a well-known phenomenon often causing threats to the success of various geotechnical projects, including tunneling, road and bridge construction, and geothermal drilling. The origin of clay-sulfate swelling is usually explained by physical swelling due to clay expansion combined with chemical swelling associated with the transformation of anhydrite (CaSO₄) into gypsum (CaSO₄∙2H₂O). The latter occurs through anhydrite dissolution and subsequent gypsum precipitation. Numerical models that simulate rock swelling must consider hydraulic, mechanical, and chemical processes. The simulation of the chemical processes is performed by solving thermodynamic equations, which usually contribute a significant portion of the overall computation time. This paper employs feed-forward neural network (FFNN) and cascade-forward neural network (CFNN) models trained with a Bayesian regularization (BR) algorithm as an alternative approach to determine the solubility of anhydrite and gypsum in the aqueous phase. The network models are developed using calcium sulfate experimental data collected from the literature. Our results indicate that the FFNN-BR is the most accurate model for the regression task. The comparison analysis with the Pitzer ion interaction model as well as previously published data-driven models shows that the FFNN-BR model is highly accurate in determining the solubility of sulfate minerals in acid and salt-containing solutions. We conclude from our results that the FFNN-BR model can be used to determine the solubility of anhydrite and gypsum needed to address typical subsurface engineering problems such as swelling of clay-sulfate rocks.