Useful parameterisations of shapes for engineering models often climb from many tens to a few hundreds design variables, potentially jeopardising usual optimisation techniques. Surrogate models are intractable when parameter numbers exceed a few tens, and gradient descent through adjoint computation in high-dimension is threatened by potential multimodality. Dimension reduction addresses these problems nicely. This paper aims to improve upon the Active Subspaces dimension reduction method by applying a combination of active subspaces in subregions of the design space, where their use of the objective function’s gradients will exploit useful local information to discover specific trends instead of trying to identify global trends over an entire design space. The partitioning of the input space is done through Gaussian Mixture Model clustering, where the authors assume the distribution of the joint inputs/outputs to be a mixture of Gaussians. This GMM clustering is used to drive a Support Vector Machines (SVM) based supervised classifier. The SVC classifier’s margin (the maximal separation distance between points of different clusters) is chosen using cross validation in order to set an optimal width for the overlapping zones. The use of overlapping increases prediction accuracy at the boundaries between clusters, an improvement over recombination methods typically found in the literature. The local expert of choice is Kriging, in part due to its built-in variance prediction.< Réduire