Automated Conceptual Design
How can we evaluate basic design configurations before metric parameters are set? When we brainstorm solutions for a problem, engineers ideate many candidate designs. How many wings should my aircraft have? Do I power this actuator with an electric motor or pressurized fluid? Engineers come up with solutions that promote a diverse pool of candidate designs to better ensure that effective, novel configurations are not missed in the design space. They then narrow the design pool to a few good candidates as best as they can without experimentation. But how do we systematically describe this process and can AI replicate or do it better than humans?
Example of 2D embeddings of different milk frother sketches using similarity judgments provided by humans (right side) versus a Large Language Model (left side).