Unique Research · Insight · November 7, 2023 · Adapted from Didem Gürdür Broo, IEEE Spectrum · Original: spectrum.ieee.org/ai-design
Author: Didem Gürdür Broo — Assistant Professor, Department of Information Technology, Uppsala University (Sweden). Leads the cyber-physical systems lab on data-centric methods for sustainable, human-centered intelligent systems (collaborative robots, AVs, smart cities). Previously Marie Skłodowska-Curie Fellow at Stanford; researcher at Cambridge.
How genAI is guiding innovative design
2023 may be remembered as genAI’s year one. Whether it makes life better remains open — but new AI tools keep shipping, and engineers gain by experimenting and folding them into design.
Examples already exist. For Aston Martin’s DBR22 concept, designers used AI integrated into Divergent Technologies’ digital 3D software to optimize rear subframe shape and layout — organic, skeletal forms via AI form exploration, then additive manufacturing. Aston Martin says weight fell while stiffness held; same design/manufacture process planned for upcoming low-volume models.
NASA research engineer Ryan McClelland calls commercial-AI-designed 3D-printed parts “evolved structures” — unfamiliar-looking yet higher structural load at lower weight than traditional components, in a fraction of the design time. AI iterates design mutations toward high performance. In such environments, co-design with genAI, high-quality structured data, and well-studied parameters clearly yields more creative and effective designs.
How genAI sparks engineering design
In January last year the author began experimenting with genAI for cyber-physical systems research (smart homes, AVs) — physical + compute components with feedback loops. Designing them needs creative collaboration between designers and engineers — slow. Could AI generators widen option space, speed iteration, or bridge disciplines?
She sought inspiration more than prescriptions. Text and music generators were fun; image generators fit best. Platforms tried: Craiyon, DALL·E 2, Midjourney, NightCafé, Stable Diffusion — best combo Midjourney + Stable Diffusion (proprietary online vs. free local open-source).
First experiment: co-design a self-sustaining jellyfish robot planned for Uppsala’s lab — nature-inspired CPS collecting microplastics as part of marine ecosystems. Usual lab loop: brainstorm → sketch → CAD → simulate → prototype → test — collaborative but experience-heavy. Vague early prompts failed; more precise prompts synced with vision; textures/materials iterated until several designs satisfied. Minutes for first results; hours to change, restate, combine successful elements.
Good prompts are specific (purpose) yet leave room for surprise — balance known wants vs. unknown-but-maybe-wanted. Unspecified attributes may be randomly assigned; include composition instructions for product design.
Controlling genAI outcomes is nearly impossible
Trying to control co-creation sometimes worked, mostly failed. Humanoid-robot “technical design” prompts often injected meaningless pseudo-letters — visual quirks from blueprint-heavy training data, not real language. Jellyfish → octopus failed oddly (engineers see them as similar) — likely training-set gaps; later octopus images slightly less crude still anatomically wrong.
Promising jellyfish designs reviewed with the team for aesthetic/functional cues transferable to physical prototypes (arched umbrella tops → protective shells; flowing tentacles → flexible manipulators). None copied directly; organic shapes sparked useful ideas and research — resisting premature lock-in. IDEO CEO Tim Brown: prototypes “slow us down so we can speed up,” avoiding costly early over-complexity and clinging to weak ideas.
Even unsuccessful genAI results can inspire
Image generators for smart-city communication complexity gave creative freedom vs. Visio/Illustrator — but poorly depicted flows and interactions; models don’t grasp context or relational meaning. After months pushing generators to limits: useful for exploration, inspiration, and quick brainstorm illustrations. Even unrealistic images push new directions; even bad flow diagrams spark productive brainstorms.
Co-creation needs persistence. Fast good results delight; specific agendas make tools hard to steer. Human control of iterations is weak; outcomes unpredictable. The author’s own thinking, expression, and taste shifted mid-process — sometimes abandoning initial specs for AI-suggested visuals; sometimes loosening constraints so the AI had more freedom. Process changes both design and designer.
Generators have updated toward more predictability — negative if you want unconventional concepts; understandable if you need control. Future: tools that execute predictably under clear constraints; image generators integrated into engineering tools and trained on their data.
Serious ethics: amplifying demographic/other biases; misinformation; privacy and IP; livelihoods of artists/writers. Transparency, oversight, accountability on data sources, generation, and downstream use are required — users must treat these issues seriously.
If used ethically, tools help engineers explore future systems, leave comfort zones, and create productive “chaos” before rigorous engineering imposes order — thinking differently, seeing connections, considering future impacts, designing innovative sustainable solutions.
7 CN images; attach on ingest.
— END —