Open Access

Carola Lenzen, Wolfgang Reif

• Assembly modeling in computer-aided design (CAD) refers to designing new products based on a collection of preexisting individual parts. To streamline this process, designers would benefit from recommendations for parts needed next, tailored to a specific extension point within the design. By their nature, assemblies can be represented as undirected graphs over parts. As parts of an assembly can be inserted in any order, we employ graph neural networks (GNNs) that are invariant to permutations. In terms of graph machine learning, the problem of localized part recommendation does not match traditional formulations such as link prediction or purely generative tasks that mostly focus on generating graphs with specific statistical properties on a macro-level. Instead, a novel approach is required that integrates the prediction of parts along with their connection to the existing graph at a specific node. In this problem setting, we investigate two distinct use cases: predicting new partsAssembly modeling in computer-aided design (CAD) refers to designing new products based on a collection of preexisting individual parts. To streamline this process, designers would benefit from recommendations for parts needed next, tailored to a specific extension point within the design. By their nature, assemblies can be represented as undirected graphs over parts. As parts of an assembly can be inserted in any order, we employ graph neural networks (GNNs) that are invariant to permutations. In terms of graph machine learning, the problem of localized part recommendation does not match traditional formulations such as link prediction or purely generative tasks that mostly focus on generating graphs with specific statistical properties on a macro-level. Instead, a novel approach is required that integrates the prediction of parts along with their connection to the existing graph at a specific node. In this problem setting, we investigate two distinct use cases: predicting new parts for a given partial design and user-selected extension point, as well as recommending both a new part and its extension point within the existing design. Our experiments indicate that our approaches significantly reduce the cognitive burden for designers: When recommending ten potential next parts, they included the needed part in up to 97.5% of cases for the first and both the part and its location in up to 92.0% for the second use case.…