An AI-Driven Design Framework for Programmable Material Systems

A team with Northwestern Engineering’s Wei Chen has developed an artificial intelligence (AI) framework that could enhance the design and functionality of programmable material systems (PMS). 

Crafted from smart materials, these structures can respond to external stimuli and transitioning between multiple states or functions. The applications of PMS are diverse and impactful, ranging from surgical robots and deployable satellites to water and energy harvesting devices. However, PMS have posed challenges because the co-design of structure, material, and external stimuli is a non-trivial task due to the extremely vast space of possible choices, high dimensionality, and a trade-off across on-demand functions. 

To combat those obstacles, the team developed a data-driven design approach based on a state-of-the-art machine learning technique called “neural operator.”  

“We have broadened the horizon of AI for design, specifically for PMS featuring complex fields of inputs and outputs, beyond the reach of conventional design,” Chen said. “This could help expedite customization and miniaturization of next-generation products and devices with built-in intelligence.”

Chen is the Wilson-Cook Professor in Engineering Design and professor and chair of mechanical engineering at the McCormick School of Engineering. Chen and her colleagues presented their research in “Deep Neural Operator Enabled Concurrent Multitask Design for Multifunctional Metamaterials under Heterogeneous Fields,” published March 26 in Advanced Optical Materials. Postdoctoral researcher Doksoo Lee was the first author.

To make this advancement, Chen and her teammates developed the Implicit Fourier neural operator technique, which can provide effective field-to-field surrogate modeling in programmable metamaterials design. With this technique, the inputs are a pair of high-dimensional structure-stimulus fields and the outputs are high-dimensional, heterogeneous physical fields that can be converted into desired functionalities.

By using neural operators along with a multiobjective optimization method that aims to find the best balance, researchers can create designs for materials that respond to external prompts instantly. This approach was successfully applied to create a type of nanoantenna array that can be programmed to do specific tasks. By tweaking the shape of the nanoantennas and how they interact with light or other energy forms, the researchers made the antenna focus energy in different patterns as needed, making it quicker and easier to design advanced materials for specific uses. 

“Our work is the first that combines the neural operator approach with design optimization for co-design of structure-stimulus in programmable metamaterials. This can significantly shorten the design cycle for achieving on-demand programmability with increased design degrees of freedom,” Chen said. “Our machine learning-assisted design framework offers intelligent ‘navigation’ in the extremely large, entangled joint design space of architecture and external stimuli. The achieved optimality, tunability, and new functionality are unattainable with the conventional optimization approach.” 

This [breakthrough] could help expedite customization and miniaturization of next-generation products and devices with built-in intelligence. Wei Chen

This study builds upon previous work by the same team, which explored the design of phase distribution (stimuli) with fixed architecture (geometry). The new framework represents a leap forward, offering access to local solution details and achieving a prediction speed that is orders of magnitude faster. This acceleration allows the use of optimization with numerous initial starting points to identify designs that represent the best possible trade-offs among various functionalities, allowing for the production of next-gen devices such as personalized wearable devices, soft robots, and smart adaptive structures.

Looking ahead, the research team plans to extend its AI-driven design approach to additional PMS applications, further exploring the potential of this technology. The focus will also include advancing the design framework to account for uncertainties, such as those arising from manufacturing imperfections, enhancing the reliability and applicability of PMS in real-world scenarios.

This work is part of the National Science Foundation Boosting Research Ideas for Transformative and Equitable Advances in Engineering (BRITE) Fellow award.