PINN-Based identification of spatially varying elastic moduli from experimental full-Field displacement data

Recent advances in imaging techniques and digital image correlation (DIC) have made it possible to acquire rich full-field displacement data, opening the door to the proper identification of material properties. Yet, conventional identification methods face severe limitations when tackling high-dimensional parameter spaces: constitutive models with large numbers of parameters or spatially varying properties. To address this challenge, we extend the physics-informed neural network (PINN)-based inverse framework into an effective methodology tailored to realistic experimental mechanics settings. The proposed approach builds on a mixed PINN formulation, where displacement and stress fields are represented by distinct neural networks (NNs), and incorporates several key innovations such as: enforcement of multiple global mechanical equilibria to exploit experimentally accessible reaction forces; Fourier features embeddings into the NNs to capture high-frequency components; finite element meshes for representing property fields; and a dedicated initialization and alternating minimization strategy ensuring convergence in high-dimensional coupled mechanical and NN parameter spaces. The study focuses on the identification of spatially distributed elastic moduli. The methodology is first validated on synthetic data, accurately recovering a complex Young’s modulus distribution and demonstrating the method’s strong ability to reduce measurement noise. It is then applied to DIC-based experimental displacements from a perforated plate loaded to failure. The approach successfully identifies homogeneous elastic constants in the initial regime, in agreement with the well-known finite element model updating method, and subsequently reconstructs Young’s modulus fields that reveal early damage localization and magnitude well before macroscopic crack initiation. To the best of our knowledge, this is the first time that such a field has been identified in experimental mechanics. The methodology proved not only to be accurate but also computationally efficient, only requiring standard computational resources.