Direct data-driven algorithms for multiscale mechanics

We propose a randomized data-driven solver for multiscale mechanics problems which improves accuracy by escaping local minima and reducing dependency on metric parameters, while requiring minimal changes relative to non-randomized solvers. We additionally develop an adaptive data-generation scheme to enrich data sets in an effective manner. This enrichment is achieved by utilizing material tangent information and an error-weighted k-means clustering algorithm. The proposed algorithms are assessed by means of three-dimensional test cases with data from a representative volume element model.