Across industry, applications involving Artificial Intelligence are shifting from task-specific to general purpose foundation models able to perform a diverse set of previously unseen functions with minimal instruction or additional training. To develop such a foundation model for engineering design, training must be completed at a meaningful scale on artifacts of prior product development, which can be multimodal and sparsely annotated. This work presents a sequence learning framework for training a foundation model on contextual relationships between function, form, and fabrication in engineering design. This learning method is demonstrated with a case study in absorbent product design.

A key challenge for implementation of a circular economy model in manufacturing systems is the functional dependence of downstream processes on upstream byproducts. Design principles provide a framework for mapping goals to solutions by decomposing complex engineering problems into structured sets of requirements to be satisfied and embodied by design parameters and process variables. Large Language Models can computationally represent such textually-described design elements to quantify interconnections between problems, solutions, and processes. We present a Functional Digital Twin concept, powered by AI language modeling and guided by principles of manufacturing systems design, to identify functionally coupled process variables in an industrial symbiosis and automatically push alerts to stakeholders in a circular manufacturing system. Changes in byproduct composition are pushed downstream, and upstream decision-makers are guided to balance satisfying their design requirements with maintaining circularity of the system. The presented method is demonstrated in a case study of bio-based absorbent materials for intended use in disposable sanitary articles developed from byproducts of the agro-food industry.