Reconstructing random heterogeneous media through differentiable optimization

Microstructure reconstruction is a key enabler of process-structure–property linkages, central topic in materials engineering. Revisiting classical optimization-based techniques, they are recognized as powerful framework to reconstruct random heterogeneous media, especially due their generality and controllability. The stochasticity the available approaches is, however, identified performance bottleneck. In this work, approached differentiable optimization problem, where error generic prescribed descriptor minimized under consideration its derivative. As an exemplary descriptor, suitable version spatial correlations formulated, along with multigrid scheme ensure scalability. applicability realized through demonstrated using wide variety achieving exact statistical equivalence errors low 0% short time. We conclude that, while still early stage development, approach has potential significantly alleviate computational effort currently associated reconstructing general media.