How Generative AI is Boosting Manufacturing Design
Faez Ahmed works toward a ‘general artificial designer’ that can learn from many different designs, transfer ideas across domains, and deliver novel and highly practical results.
“In the next few years, the way we design and manufacture products is going to dramatically change,” says Faez Ahmed. According to Ahmed, design tools that have remained largely unchanged for decades will soon undergo a complete overhaul to incorporate large language models and other AI technologies. These enhancements will make the tools far easier to use, and expand the ways in which people currently design products.
Ahmed, the ABS career development assistant professor in mechanical engineering, studies how generative AI (GenAI) can help humans create better designs faster for engineering products.
Ahmed points out that GenAI algorithms can enhance the efficacy of classic engineering simulations and design optimization techniques, which, despite their power, are often constrained by time-consuming problem formulation and iterative processes. In contrast, “GenAI models learn from existing designs, and they can give you a very good option that works,” Ahmed says. “Taking the power of optimization tools and marrying it with GenAI models, you can get the best of both worlds.”
Taking the power of optimization tools and marrying it with GenAI models, you can get the best of both worlds
One of his favorite examples is a collaboration with BikeCAD, which has empowered engineers and enthusiasts to invent thousands of computer-aided design (CAD) bicycle designs. Ahmed and his colleagues trained and polished GenAI models for BikeCAD's rich set of data. Some of their preliminary AI models are now deployed with the BikeCAD software, which can aid in generating unique, creative, and high-quality bicycle designs.
Breaking barriers with engineering data “GenAI is everywhere right now in the text and image worlds,” Ahmed says. “But GenAI's impact on engineering, and specifically design and manufacturing, is still limited. There are three key challenges.”
The first key challenge is precision. “How can we create precision generative AI models, which give the right answer accurately?” he asks. “Our work has shown that there are ways where you can combine generative AI with optimization methods to do that.”
The second challenge is multi-modality—enabling AI models to understand many different types of CAD, text, image and other forms of data.
“Third, which is the most interesting challenge, is innovation,” Ahmed says. GenAI software mimics the data that it trains on, which works well for generating human faces and many other types of image and textual data. But engineers want to find innovative solutions.
Additionally, another major obstacle is the lack of huge volumes of data in the design and manufacturing world for training AI. “Many people don't realize that data, not the models, is central to GenAI's effectiveness,” Ahmed remarks. “A key challenge which faces our domain is that we are not blessed with data.”
That’s one reason he strongly encourages companies to think carefully about how their domain data can help them by leveraging powerful AI tools. Even data sets that have faulty data or are small in scale can provide useful data.
GenAI modeling goes wide The process of making a product has many commonalities across many domains. “You start with conceptual-stage sketches, you refine the design, you create prototypes and you do simulations,” Ahmed says. “That process is everywhere, whether you're designing a paperclip, a ship, an aircraft component, or a bicycle.”
His group looks for applications that might help in developing a “general artificial designer” that can learn from many different designs, transfer ideas across domains, and deliver novel and highly practical results.
One notable project involves a sketch-to-prototype GenAI application that began "almost accidentally" when Ahmed and his students considered the classes where undergraduates sketch designs. Conceptual sketches created by students are hard to translate to engineering tools. Could GenAI go beyond sketches to generate more realistic and useful prototypes?
The researchers found that large language model (LLM) and other GenAI software could accelerate the sketch-editing process. They showed that if designers could interpret and modify their sketches with an LLM model, and use those results to produce three-dimensional models, the quality of the final product could significantly improve, "which was unexpected," Ahmed says. Equally important, editing sketches via LLM software allowed designers to explore their new ideas more quickly.
Another effort took on the design of ships, which is an extremely tricky process with complex trade-offs. “The shipping industry is poised for major changes due to new carbon emission regulations,” Ahmed observes. “They are rethinking everything, so they are very excited about what role AI plays in designing the ships for the future.”
In one project, he and his collaborators explored improved methods for designing ship hulls using GenAI software. The team analyzed existing hulls, generated random designs for training GenAI models, and then trained these models to achieve objectives like minimizing friction during operation. This approach resulted in innovative, high-performing designs that could potentially reduce ship operating costs and improve sustainability.
Combining physical and digital advances As an undergraduate at the Indian Institute of Technology in Kanpur, Ahmed studied how autonomous navigation vehicles could find routes more efficiently. Solving such design problems “has ties to both optimization and machine learning,” he says. “That's how my research trajectory started going from physical robots to the digital.”
He gained more motivation for this research when he later worked for railways at a remote mining site in Western Australia. (Figuring ways to guard against crashes of giant trains that could stretch 1.5 miles long “was a very exciting thing for a mechanical engineer,” he recalls.)
While completing a doctorate in mechanical engineering at the University of Maryland in College Park, Ahmed was recruited by MIT. “The faculty, the students, the research staff, and everyone here are really excited and passionate about solving very complex problems in mechanical engineering,” he says. Moreover, MIT's highly interdisciplinary research environment is a huge help in looking at the problems from a holistic perspective.
Innovating with industry Corporate collaborations are essential for progress in his research. “We are very excited about frequent collaborations with industry,” Ahmed says. Participating in MIT programs with IBM and Toyota Research, he is always looking for “fun people and fun ideas” to work on. He is the American Bureau of Shipping (ABS) Career Development Chair and has received the 3M non-tenured faculty award and the Google Research Scholar award.
Ahmed highlights the potential of GenAI to democratize engineering design, allowing anyone, even those without engineering training, to contribute valuable ideas and expertise. "We can tap into the creative potential of many more people, rather than relying solely on the design departments of top companies," he says.
“It's a very exciting time to be a mechanical engineer,” Ahmed remarks. "It's like you get to choose which major issue you want to tackle. There are fascinating parallels between various problems that initially seem very different."
“While many engineering students today concentrate solely on the digital world if you apply digital engineering, machine learning, and AI to solve problems in the physical world, you can have much more impact on significant challenges,” Ahmed adds. “We can explore how AI can enhance products, making them better, faster, and more innovative. Additionally, sustainability and climate change are critical issues where AI and design can play a crucial role.”