Skip to main content

Illinois Computes Enables Novel Engineering AI Research


An image of Delta cabinets in the NPCF machine room.

The NCSA Delta System.

Researchers from the National Center for Supercomputing Applications and The Grainger College of Engineering at the University of Illinois Urbana-Champaign (UIUC) recently published in two prestigious engineering journals their work on novel deep operator network (DeepONet) implementations to more accurately and efficiently predict stress responses in complex geometries.

A multidisciplinary team from NCSA and the Department of Mechanical Science and Engineering (MechSE) completed neural network training and inference on NCSA’s Delta system through Illinois Computes, a program offering computing and data storage resources, technical expertise and support services to researchers from all domains across the UIUC campus at little to no cost.

“The Illinois Computes resources have been extremely valuable for my research,” said Iwona Jasiuk, professor in MechSE and Richard W. Kritzer Faculty Scholar. “This comprehensive program has provided all the needed resources to advance my research in machine learning and computational mechanics at low cost. The technical support has been impressive and the application process was straightforward. This program facilitated the interdisciplinary collaboration with the NCSA researchers. We are jointly addressing open scientific questions and pressing technological problems while advancing computational tools.”

Delta is NCSA’s premier system and the most performant GPU-computing resource in the U.S. National Science Foundation’s portfolio, making it ideal for the artificial intelligence and machine learning applications utilized for this DeepONet research.

“Delta, with its unique computational capabilities, is instrumental in our research,” said Seid Koric, technical associate director for the Research Consulting Directorate at NCSA and associate research professor at MechSE. “We used its high-throughput and parallel CPU capabilities to efficiently generate thousands of training data samples from nonlinear finite element analysis with Abaqus software with NCSA’s academic license. Then, we used Delta’s A100 Nvidia GPU nodes to train deep and challenging artificial neural networks in DeepONets.”

Appearing in “Computer Methods in Applied Mechanics and Engineering,” the DeepONet uses a residual U-Net (ResUNet) in the trunk to encode complex input geometries and a fully connected branch network encodes the parametric loads. For learning complex, nonlinear operators, the DeepONet is a recently proposed neural network architecture that has shown tremendous success in approximating complex physics operators such as heat diffusion, plastic deformation, multi-scale analysis, crack propagation, Darcy flow over complex domains and engine combustion. The novel DeepONet was tested against two baseline models and significantly outperformed both as it is more memory efficient and allows greater flexibility with framework architecture modifications.

Stress solution comparison, DeepONet prediction vs. Material Nonlinear (Plastic) Finite Element (FE) solution, with trained DeepONet accurately inferring solution more than two orders of magnitude faster than FE for an arbitrary loading history.

This work marks the first time a ResUNet is used as the trunk network in the DeepONet architecture and the first time that DeepONet solves problems with complex, varying input geometries under parametric loads and elastoplastic material behavior.

In the second paper recently published in “Engineering Applications of Artificial Intelligence,” the team devised another novel implementation of DeepONet, called S-DeepONet, based on advanced sequential learning methods. Trained on Delta, S-DeepONet provided increased accuracy of entire multi-physics solution fields under arbitrary thermal and mechanical loading histories and did so up to 10,000 times faster than classical numerical methods. 

This innovative research, which is a continuation of previous work by Koric and Abbueida on DeepONets, opens the door for drastically accelerating future research discovery, modeling, design and optimization of advanced and additive manufacturing processes and the creation of digital twins for many applications in science and engineering. 

“Additive manufacturing is a revolutionary manufacturing technique that opens nearly unlimited possibilities for its implementation,” said Jasiuk. “DeepONet serves as a powerful and rapid computational tool, which can simulate the additive manufacturing process at various spatial and temporal scales. Such simulations are needed for deeper understanding of the additive manufacturing process and its implementation and monitoring.”

NCSA Research Scientist Diab Abueidda and engineering graduate students Junyan He, Shashank Kushwaha and Jaewan Park co-authored the paper with Jasiuk and Koric.

“The experience has been enriching and educational, highlighting the collaborative synergy of multidisciplinary expertise and cutting-edge technology,” said Abueidda. “Utilizing Illinois Computes and the Delta system has provided an exceptional platform to explore the frontiers of artificial intelligence in engineering, enabling us to conduct innovative research with significant impact.”

“Novel AI/ML research cannot be made possible without the support of state-of-the-art GPU computing resources, which were kindly provided to us via the Delta project,” said He. “Thanks to the computational resource, we were able to quickly iterate on different architectures and improve model performance, which greatly accelerated our pace of research. The collaboration between NCSA and MechSE brings talents from both sides together to effectively tackle challenging problems in mechanics via advanced algorithms in machine learning, and I am glad that I participated and contributed to this partnership.”

“Participating in this project and engaging with the Delta system has been a profoundly educational and academically enriching experience. The availability of various finite element software, like Abaqus, has helped us quickly generate data for neural network training,” Kushwaha said. “The opportunity to utilize advanced GPU-computing resources under the Delta project has expedited our research process and enabled us to push the boundaries of traditional engineering methods. This collaboration between NCSA and MechSE has reinforced the importance of integrating different technological domains to address complex problems in mechanics and engineering.”

It’s been an exhilarating experience diving into the world of novel AI and DeepONet research, especially with the incredible power of the NCSA’s supercomputer Delta at my disposal thanks to the Illinois Computes program. The immense computational capabilities of Delta have opened new horizons in my research, allowing me to explore complex neural network architectures and deep learning algorithms with unprecedented efficiency and scale. The versatility of DeepONet in capturing intricate functional relationships within data is particularly fascinating, as it paves the way for breakthroughs in predictive modeling and data-driven discovery. It’s a remarkable period of learning and innovation, where the convergence of advanced AI methodologies and powerful computing infrastructure is leading to transformative advancements in scientific research.

Jaewan Park, Engineering Graduate Student

Illinois Computes proved instrumental to this project, providing computing time on Delta and advancing the team’s research.

“Illinois Computes is much more than the utilization of high-performance computing, data and other great cyberinfrastructure hosted at NCSA,” Koric said. “It’s an outstanding opportunity to promote and support multidisciplinary collaboration between researchers at NCSA and UIUC, who are jointly working on solutions for grand challenges in science and society.”

Illinois Computes offers computing and data storage resources, technical expertise and support services to researchers from all domains across the UIUC campus at little to no cost. Through the campus-funded program, NCSA will learn what additional assets are needed to fulfill the computing demands of the university and adjust the cyberinfrastructure strategy while continuing to make access to systems, interdisciplinary and technical knowledge, and support infrastructure easy to obtain. Illinois Computes removes barriers for all Illinois researchers – especially those typically underserved – to access NCSA’s growing assemblage of research computing tools and world-class staff, furthering their innovative and novel work while ensuring NCSA is a leader in the global research community.

Check out the Illinois Computes website for more information or to get involved.

Back to top