Extraction of Implicit Knowledge and Optimization

Those last years have seen a growing maturity in optimization tools, particularly topological optimization [1]. Nevertheless, those methods do not take the accumulated knowledge of industrial manufacturers into account. Moreover, they face substantial limitations to include all the various constraints that govern the realization of industrial products. We wish to ally discovery tools from IA and optimization ones in this project. A key question is that the amount of design of a given industry is very low and thus it is not possible to use deep-learning approaches for this purpose. Therefore, among the possible tools we have elected those allowing to infer, if possible, a reduced manifold of the existing design. If the latter is of low dimension optimization my then be performed in this manifold. The second question is what type of descriptors and classification are the most appropriate to perform the dimension reduction and to interpolate properly within the existing design. Those tools combine reduced modelling and manifold learning approaches (like Locally Linear Embedding) [1], geometrical characterization (Level set [2]), Topological Data Characterization [3] and optimal transport [4]. Those tools are today of easy access through dedicated libraries. Those basic concepts will be illustrated with 2D examples mimicking bumpers. At the present stage, it seems that we are capable of defining the geometrical manifold of bumpers and interpolate following the manifold. We are looking to incorporate mechanical information such as von Mises maps to enrich the description of the manifold. Our present focus, which we hope to be able to illustrate during the conference, is to make use of the enriched manifold to be able to optimize along the variety of “validated design’, that is, to create know-how optimization tools.