A Mixture of Physics-Informed Expert system for robust prediction of concrete failure behavior under high impact loads

Predicting concrete failure under high-impact loads is critical for designing protective structures. Machine learning models can provide high predictive accuracy, but they often lack robustness when applied to out-of-distribution data. Traditional physics-based methods cannot fully resolve this limitation because concrete failure under impact loading is highly nonlinear and difficult to describe with complete physical accuracy. To address this challenge, this study proposes a Mixture of Physics-Informed Expert system, an artificial intelligence and deep-learning framework for predicting concrete failure behavior under high-impact loading. The proposed system uses physics-based gating logic and a dynamic regularization strategy to improve prediction reliability when the input conditions differ from the training data. The system was evaluated for penetration depth prediction and achieved a coefficient of determination of 0.9884. It also reduced the mean absolute error by 9.7% compared with a standard neural network and by 68.0% compared with common empirical formulas. These results show that the proposed artificial intelligence approach can balance data-driven accuracy with physics-based reliability for robust prediction of concrete failure behavior in critical engineering applications.