An Inverse Parameter Identification in Finite Element Problems Using Machine Learning-Aided Optimization Framework

The ability of finite element analysis to produce high fidelity results is greatly dependent on quality of constitutive model and the accuracy of their parameters. As such, the calibration of phenomenological constitutive models to replicate real-world behaviors has remained a focal point of many research works. A new inverse identification approach combining numerical-experimental methods and data-driven techniques to characterize the nonlinear response of materials using a single experiment is proposed. This approach integrates finite element analysis, optimization methods and machine learning techniques, such as Artificial Neural Networks and Support Vector Regression, to accurately determine model parameters while significantly reducing computational time. This approach can be used to characterize a wide range of models irrespective of the number of parameters involved. A detailed flowchart of the methodology focusing on its implementation aspects is provided and its each module is explained. The proposed model calibration approach successfully identified eight parameters for a cohesive zone model implemented in user element subroutine (UEL), four parameters for a hardening model implemented in user material subroutine (UMAT), and five parameters for a Johnson–Cook plasticity model. In all cases, this method achieved an excellent fit between the simulation and experimental results. Moreover, it demonstrated a significant improvement in efficiency, being 2–3 times faster than traditional optimization algorithms in determining optimal parameters. Based on the presented investigations, the proposed machine learning-based inverse method can significantly accelerate the parameter identification procedure and can be extended to a wide range of material models.