MechSE undergrad first author of breakthrough paper on deep learning in complex materials design
Metamaterials are artificially engineered structures with unique properties unavailable in naturally occurring materials such as wood and stone. Their value lies within their engineering potential, which accounts for architecture, constituent materials, and volume fraction to achieve the most optimal design for specific structures and spaces. The key to determining the best metamaterials for a given design space is topology optimization (TO), a mathematical method that optimizes material layout by considering load, boundary conditions, and constraints. TO focuses on material functionality and efficiency, unconstrained by traditional aesthetics and design. The sheer amount of data this method requires is nearly impossible to analyze, calculate, and process without the involvement of high-performance computing, which is not always within reach for everyone.
Access barriers include the high cost of purchasing, developing, maintaining the systems, the specialized software and technical expertise necessary to operate it, and the increasing demand further restricting its availability. Consequently, this also creates access barriers to topology optimization-based material design, limiting the development of metamaterials and architectural innovation.
Researchers and scientists from the National Center for Supercomputing Applications and its Center for Artificial Intelligence Innovation (CAII), and the Grainger College of Engineering at the University of Illinois Urbana-Champaign turned to artificial intelligence, collaboration, and deep learning to address these gaps. The team—MechSE undergraduate Hunter Kollmann, NCSA Postdoctoral Research Assistant Diab Abueidda, NCSA Research Scientist Erman Guleyruz, CAII affiliate Professor Nahil Sobh, and NCSA Technical Assistant Director and MechSE Research Associate Professor Seid Koric—recently published a paper about their research in ScienceDirect.
Leveraging CAII and NCSA Industry's iForge cluster, the team implemented deep learning to train a convolutional neural network model. "Solving a numerical topology optimization problem is extremely computationally expensive due to the large number of iterations needed and usually requires high-performance computing," said Kollmann. "Our developed CNN model can be used to reduce the computational cost of developing architectured materials, so not only is a high-performance computing system no longer required, but the results will also be available near instantly."
Koric explained further, "We can transfer the trained learnable parameters (weights and biases) to any low-end computing platform, such as laptops. In doing so, optimal solutions for any variation of input parameters can be found on such, instantly. We believe that similar AI-based surrogate data-driven models will pave the way for remarkably efficient high-fidelity design and modeling in the future—particularly for architectured and bio-inspired metamaterials."
Abueidda was the alpha and omega of this effort, envisioning this work as a student project for Kollmann. However, it quickly transitioned into cutting-edge research nominated for a 2020 HPCWire Readers’ Choice Award for Best Use of Artificial Intelligence.
"This collaborative research revolves around the confluence of finite element analysis, mathematical optimization, and deep learning. Such interdisciplinary research aims at accelerating the design process of next-generation materials," said Abueidda. A recent MechSE PhD graduate, he is excited to contribute to Kollmann's academic development and success. "It was a great experience to expose a mechanical engineering student to the recent deep learning approaches, as this might open the door for him and encourage his peers studying mechanical engineering for the broader adoption of data-driven models."
Working alongside mentors Abueidda and Koric for two years in NCSA's Industrial Application Domain, Kollmann heavily contributed to this CNN model's culmination. Grateful for this opportunity, he shares his experiences and takeaways, "What they gave me is priceless. Writing a research journal paper is very different from other forms of writing I've done. Being mentored by research scientists allowed me to ask questions and learn from them." The project required Kollmann to develop a new independent learning style to find solutions. "I did not have lecture notes to reference and could not conduct an internet search for a simple answer. Instead, I had to find and read research papers regarding the topic, which can be a lot harder to understand and requires more reading time. This is an invaluable skill to harness as a student, even post-grad."
To reward his dedication, hard work, academic enthusiasm, and zeal, Abueidda and Koric decided to bestow the honor of being the research publication's first author to Kollmann, a rare and gracious gesture in today's academic world. "For Hunter, I hope this research paper, published in the top journal, will whet his appetite for graduate study—where he certainly belongs," said Koric.
Elated, Kollmann grasps this accolade's importance and significance, "Being named first-author was extremely generous and meant a lot to me. The chance to work on and publish a research paper in a journal is an amazing opportunity in itself, let alone be named the first author of it."
Adding yet another success story to NCSA's extensive history of being a central hub that facilitates and values collaboration, innovation, and scholarship. CAII Director Eliu Huerta, a supporter of this project, understands the importance of mentorship through his involvement in various NCSA programs such as SPIN and the Gravity Group. "This is a great example of the world-class mentorship and interdisciplinary research that takes place at NCSA," says Huerta. "Seid's strong track record as a mentor and researcher is a great asset to NCSA and the activities spearheaded by the NCSA Center for Artificial Intelligence Innovation."
This project is not the first time that Abueidda and Koric's research intersected with domains such as artificial intelligence, deep learning, and high-performance computing (HPC). They recently published another paper outlining how artificial intelligence can learn and predict complex material behavior, a potential game-changer in computational mechanics.
The confluence of AI, HPC, and Big Data already plays a massive role in the technology that we interact with daily, as highlighted last year in an article written about Koric's research. Koric explains further, "It is said that 'artificial intelligence is the new electricity and data is the new oil.' AI is ubiquitous already in our lives: Image and voice recognition, language translation, autonomous cars, medical diagnosis, detecting fraudulent transactions in finance, etc.—to name a few, and they all learn from a large amount of collected data."
There are still plenty of opportunities to do more, and NCSA continues to accelerate AI-research in academia and industry. "AI may also significantly help modeling and simulation engineering design and optimization in the future. We can already generate as much training data as we want from numerical modeling on HPC nowadays. Along with our academic and industrial collaborators worldwide, we at NCSA have started developing and training innovative deep learning methods to do some difficult, time, and computationally consuming modeling operations very efficiently," noted Koric.