Learning metal plasticity constitutive models via laboratory-based X-ray measurements

Abstract

Data-rich mechanics has the potential to redefine constitutive material modelling, yet its advancement has been fundamentally limited by the absence of large-scale, high-fidelity experimental datasets capable of capturing complex multiaxial behaviour. Here, we introduce a complete experimental–computational framework that unifies large-scale laboratory dataset generation with recurrent neural-network modelling of metal plasticity. At the core of this framework is a novel laboratory-scale experimental platform that integrates energy dispersive X-ray diffraction — previously confined to scarce and costly synchrotron facilities — with optical digital image correlation to simultaneously measure diverse local multiaxial stress–strain responses from a single geometrically complex specimen. This approach enables the rapid generation of rich, high-dimensional datasets with unprecedented accessibility and throughput. Using these datasets, we demonstrate that a mechanics-constrained neural operator can accurately reconstruct highly complex yield surfaces, capturing phenomena such as tension–compression asymmetry and previously inaccessible multiaxial strain-hardening behaviour. By bridging advanced experimentation, high-throughput data generation, and transfer learning approaches, this work establishes a scalable pipeline for next-generation constitutive modelling and lays a foundation for data-rich material mechanics.