Dynamical modelling of structural beams under a digital twin framework with physics-informed neural networks

Digital Twins (DT) are used in many fields, as well as to monitor the condition of structural parts according to the loads they are exposed to during operation. In the process of creating a digital copy of the physical product, Finite Element Method (FEM) is often involved. Although the complexity of the part geometry and mesh details are important in FEM applications, they affect the analysis time and thus the generation time of DT. Creating a digital copy with high-fidelity and low latency makes it possible to employ Artificial Intelligence (AI), which has found its place in many fields. In this study, we propose a Physics-Informed Neural Network (PINN) structure and DT framework that can perform high-fidelity modeling with fewer data by using the Partial Differential Equations (PDE) of the problem during the generation of the DT of structural beams. During the studies, synthetic data are generated with the help of analytical equations of the problem. Then, the generated data are fed into the PINN model to solve the forward problem and the inverse problem is solved to estimate the system parameters. A self-adaptive weight update method with a Neural Tangent Kernel (NTK) is applied to deal with the spectral bias phenomenon in high-frequency learning. The PINN estimates are compared with the Finite Element Analysis (FEA) outputs to prove the accuracy of the developed method. After the studies with synthetic data, a study is carried out with real data collected from the experimental setup, and the potential of PINNs to approximate the solution of nonlinear higher-order PDEs and parameter identification is demonstrated.