A scalable composite Bayesian optimisation framework for material calibration using full-field measurements

Full-field measurements provide rich spatial information on heterogeneous deformation processes and have become a central ingredient in modern material parameter identification. This work develops a scalable composite Bayesian optimisation framework for response-based calibration of parametric material models using high-dimensional full-field data and expensive forward solvers. The main contribution is the extension of composite Bayesian optimisation to full-field response data, where the latent response space may contain thousands of spatial and temporal observations. To make this setting computationally tractable, the framework combines a reduced latent-space representation with efficient interpolation and data-reduction strategies, including thread-safe caching of Delaunay triangulations, localised interpolation schemes, and simplex-based decimation of reference responses. The approach is assessed on numerical and experimental calibration problems involving global and full-field response quantities, including linear elastic, elasto-plastic and nonlinear visco-elastic/visco-plastic examples. In the deterministic benchmarks considered, the composite formulation generally improves convergence behaviour and reduces parameter dispersion relative to standard Bayesian optimisation under the same evaluation budget. These results indicate that composite Bayesian optimisation provides an efficient and flexible route for exploiting high-dimensional full-field measurements in non-intrusive material calibration workflows involving computationally expensive constitutive simulations.