Reinforcement learning (RL) methods can be used for the automation of design adaptations due to their inherent iterative nature. However, they present difficulties with regard to the interpretability of the results. One reason for this is that RL methods are integrated into the product development process as black boxes. As a result, the use of these methods in safety-critical areas such as mechanical design is only possible to a limited extent due to the lack of traceability of the design decisions made by the algorithm. In order to increase the usability of RL-based tools supporting the product developers, the interpretability of the RL-based results must be improved. Therefore, this paper proposes a concept for increasing the traceability of RL-based tools for the automation of design adaptations. Traceability is to be improved by imitating the decision-making process of experts. To this end, the concept is presented using a design adaptation of fibre-plastic composite (FRP) components.

Products are developed with the objective to fulfil certain functions. Despite

these functions are rarely modeled digitally, there are implicit descriptions of them within existing product data, that can be retrieved by suited algorithms. Approaches like these are analyzed in the means of this work. This research provides a literature study on the field of automated functional modeling (AFM) in engineering design. The literature databases Scopus and Web of Science are searched systematically for relevant publications between 2010 and 2025. 19 approaches for AFM are identified and structured with regard to their motivation for AFM, data input, data output and data processing. Classification criteria for labeling these four categories are defined inductively while analyzing the publications. The AFM approaches are then structured as matrices according to their properties. Main insights are, that most of the AFM approaches operate on textual or CAD design data and that machine learning seems to have a huge potential for AFM.

Dieser Beitrag untersucht die Bedeutung von unstrukturierten und heterogenen Daten für die Entscheidungsfindung im Produktentwicklungsprozess bei technischen Änderungen. Entscheidungen müssen auf validen Informationen und dem Einbezug von explizitem und implizitem Wissen basieren, wobei eine nachvollziehbare Dokumentation unerlässlich ist. Eine Ist-Stand-Analyse eines Automobilzulieferes (Tier 1) zeigt, dass wertvolles implizites Wissen insbesondere in unstrukturiertem Daten enthalten ist. Diese sind jedoch schwer zu analysieren und zu validieren, da die Daten nicht durchgängig und in unterschiedlicher Qualität vorliegen. Es wird vorgeschlagen, technologische Ansätze zur Wissensmodellierung anzuwenden, um das Wissen für zukünftige Projekte nutzbar zu machen. Herausforderungen bestehen in der Sicherstellung der Datenqualität und -validität sowie der Integration neuer Technologien. Zukünftige Forschungsaktivitäten sollen sich daher der Analyse technologischer Verfahren widmen, um zweckdienliche Lösungen, ggf. durch gezielte Kombination verschiedener Technologien, zu untersuchen und zu definieren.