An Efficient PINN–Based Calibration Method for Mesoscale Peridynamic Concrete Models

Mesoscale models are crucial for the refined analysis of material damage behaviors. However, it remains a challenging task to calibrate a mesoscale model so as to accurately simulate the mechanical behaviors (MBs) of macroscale structural components. The models may be nonlinear, involve numerous material parameters (MPs), and be large-scale. In addition, solutions to inverse problems may lack accuracy or be nonunique. A recent emerging method, physics-informed neural network (PINN), combines deep learning with physical laws to solve complex problems and significantly reduce computational costs. This paper presents an effective PINN approach for mesoscale model calibration. The approach establishes a relationship between the MPs of a mesoscale model and the MBs of structural components using PINN, with constraints based on known physical relationships. Both forward PINN (MPs as inputs and MBs as outputs) and reverse PINN (swapping inputs and outputs) models are used. Calibration is achieved efficiently by combining the forward PINN model with an optimization algorithm or directly using the reverse PINN model. Validation is performed using a mesoscale concrete model in peridynamics (PDs). The relationship between the elastic modulus of bonds in PD and MBs of components is constrained by physical laws. The datasets are generated through OpenSees analysis. The PINN method demonstrates its effectiveness, particularly with the reverse model, which is both efficient and accurate.