Abstract
We introduce the first Automatically Differentiable Model Updating (ADiMU)framework that finds any history-dependent material model from full-fielddisplacement and global force data (global, indirect discovery) or fromstrain-stress data (local, direct discovery). We show that ADiMU can updateconventional (physics-based), neural network (data-driven), and hybrid materialmodels. Moreover, this framework requires no fine-tuning of hyperparameters oradditional quantities beyond those inherent to the user-selected material modelarchitecture and optimizer. The robustness and versatility of ADiMU isextensively exemplified by updating different models spanning tens to millionsof parameters, in both local and global discovery settings. Relying on fullydifferentiable code, the algorithmic implementation leverages vectorizing mapsthat enable history-dependent automatic differentiation via efficient batchedexecution of shared computation graphs. This contribution also aims tofacilitate the integration, evaluation and application of future material modelarchitectures by openly supporting the research community. Therefore, ADiMU isreleased as an open-source computational tool, integrated into a carefullydesigned and documented software named HookeAI.