An approach to use generic data sets for neural networks in product designs through geometric abstraction via primitives

Generative design suggestions and topology optimizations can help to reduce iterative process loops between calculation and design departments during product development processes. However, precise topology optimizations are computationally intensive, while generative designs benefit from swift suggestions to address design problems efficiently. Using artificial neural networks (ANN) can address this contrast of pre-defined aims by predicting topology-optimized designs, thereby combining both advantageous features. However, a challenge in Mass Customization is, that ANN are usually trained on specific geometries, making transfer to other applications impractical or requiring the creation of new datasets, which is economically unfeasible. Authors have already demonstrated a solution in other publications: dividing a geometry into geometric primitives like cuboids to perform abstraction. An ANN can then be trained to recognize optimized cuboids, which can be assembled back into a complete geometry, comparable to the finite element methods, which divide geometries of parts in finite elements enable mechanical property calculation. This publication aims to illustrate the steps of the approach in which the complete geometry of a part is segmented into these primitives, and the benefits obtained. Various methods will be explored, including automated workflows on modern low-code platforms, to enable generalized use.