Empirical physics-informed neural networks for prediction of concrete strength using nondestructive testing

Abstract

Predicting the compressive strength of concrete in built structures is crucial for assessing structural safety, addressing damage, adapting to regulatory changes, determining repair needs, and evaluating the strength of components for reuse in sustainability. Non-destructive testing (NDT) has been effectively used for decades, initially through empirical equations and, more recently, through machine learning (ML) models to predict the compressive strength of concrete in existing structures. Both traditional empirical equations and ML models have demonstrated promising results, but each has inherent limitations. This study introduces Empirical Physics-Informed Neural Networks (EMP-PINNs), which integrate empirical equations with artificial neural networks (ANNs) to combine the advantages of both methods, presenting an innovative fitting algorithm. First, a comprehensive dataset was generated using Generative Adversarial Networks (GANs) to enhance the training of machine learning models. This dataset includes NDT values such as rebound number and ultrasonic pulse velocity, with concrete strength as the target variable. Next, a regression equation was developed to serve as a physics-informed constraint within the EMP-PINN model. Finally, the empirical equation was combined with ANN to form the EMP-PINN model. The developed EMP-PINN model demonstrated strong convergence behavior, with a high correlation coefficient (R² = 0.92) and low prediction errors (MAE = 2.27, MSE = 9.29). Compared to traditional ANN, EMP-PINN showed superior generalization, particularly for extreme or unseen values. When applied to a real-world structure, the model achieved an average absolute prediction error of approximately 5.6%, validating its practical reliability. This Empirical Physics-Informed Neural Network paves the way for the broader application of physics-informed neural networks in engineering domains, where governing systems are often based on empirical equations rather than purely physical ordinary differential equations.