Physics informed neural network with multiple subnetworks for predicting rheological behavior of cementitious materials

Traditional Physics-Informed Neural Networks (PINNs) utilize a fully-connected architecture, causing shared parameters among outputs and subsequent accuracy degradation. We proposed a MultiSubPINN model to address this by deploying independent networks for distinct predictions, minimizing output interference and preserving individual output optimizations. Our results demonstrate that MultiSubPINN significantly outperforms conventional PINN models in solving complex partial differential equations (PDEs). Notably, MultiSubPINN achieved a reduction in Mean Square Error (MSE) by 32.7%, 9.9%, and 52.8% for shear stress, viscosity, and structure parameter predictions, respectively. Moreover, we introduce a novel physics-based sampling strategy that leverages residual distribution and physics principles to further refine MultiSubPINN’s performance. This approach markedly surpasses random sampling strategies, reducing MSE by approximately 97.7%, 90.2%, and 38.3% for the aforementioned parameters. Importantly, the enhancements in prediction accuracy are achieved with a marginal increase (0.1%) in training time, underscoring MultiSubPINN’s efficacy and potential for broader application in material simulation domains. A new AI method predicts how cement paste flows and changes under stress. This model is more accurate than traditional approaches while still efficient. Findings can guide better design of sustainable and durable concrete materials.