Physics-based self-adaptive algorithm for estimating the long-term performance of concrete shrinkage

Drying shrinkage of concrete, a complex phenomenon susceptible to humidity diffusion out of concrete, represents one of the critical reasons for the structural deterioration that always affects the long-term service performance of concrete. Although various methods have been developed to forecast concrete shrinkage, more reliable and accurate prediction methods are required. This study proposes a self-adaptive physics-informed neural network (SA-PINN) for modeling the mechanism of humidity diffusion and predicting the long-term performance of concrete shrinkage. The model was enriched with a Gaussian probabilistic approach to accurately adjust the self-adaptive loss function. Specifically, the model exploits the feature settings in the neural network and leverages the loss of physical constraints as a way to fine-tune the gradient descent algorithm, which can produce more accurate results than traditional methods. The performance of the SA-PINN model is compared to that of a physics-informed network and a purely data-driven method. The results revealed that the proposed model could reasonably simulate the humidity diffusion behavior and significantly improve the prediction precision of shrinkage in concrete. Model testing provided the highest correlation coefficient (0.954), lowest mean square error (0.266), and lowest mean absolute percentage error (0.035), indicating the accuracy of this model in predicting the drying shrinkage of concrete.