University of Illinois at Urbana-Champaign
The Grainger College of Engineering
Mechanical Science & Engineering

Koric: AI successfully learns and predicts complex material behavior

9/16/2020

Seid Koric
Seid Koric, MechSE Research Associate Professor and NSCA Technical Associate Director

MechSE’s Seid Koric has published research outlining how artificial intelligence can learn and predict complex material behavior, which could be a game-changer in computational mechanics.

Their paper, “Deep learning for plasticity and thermo-viscoplasticity,” was published in the International Journal of Plasticity, the highest-impact journal in computational mechanics. 

Most engineering materials experience plastic behavior beyond the yield point and on higher-temperature temperature-dependent thermo-visco-plastic behavior. The complex response of these ubiquitous materials depends on the loading, rate, and temperature histories. Currently, various processes with these materials are widely computationally modeled in manufacturing, metallurgy, aerospace, automobile, nuclear power, life science, chemical, and many other industries. However, this requires sophisticated software and hardware resources.  

Convolutional neural networks have been proven successful in many fields, particularly in image recognition. Temporal convolutional networks (TCN) are based one-dimensional convolutional layers (Conv1D), sliding along time, and mapping sequences of an arbitrary length to output sequences of the same length. TCN does not have recurrent and data dependent calculations, and is computational and memory efficient.
Convolutional neural networks have been proven successful in many fields, particularly in image recognition. Temporal convolutional networks (TCN) are based one-dimensional convolutional layers (Conv1D), sliding along time, and mapping sequences of an arbitrary length to output sequences of the same length. TCN does not have recurrent and data dependent calculations, and is computational and memory efficient.

The authors have demonstrated how the innovative sequence deep learning methods can learn from modeling data and correctly reproduce the complex history—as well as time- and temperature-dependent phenomenon in each test example. Each of these test samples has never been seen before by the artificial neural networks. This research marks the first time the researchers applied and validated that artificial intelligence can learn and reproduce this complex material behavior successfully. Moreover, once the sufficient training data is generated and trained adequately on a high-end computing system, this material behavior can be inferenced accurately and almost instantly on any low-end computers such as a laptop and without any modeling software. The findings ensure that this and other similar AI-based data-driven methods essential for quick modeling of complex material behavior in the future.  

“To perform this challenging interdisciplinary research, our team engaged with experts from the University of Illinois as well as the innovative high-performance computing (HPC) resources and support at the National Center for Supercomputing Applications and the Center for Artificial Intelligence Innovation. We worked closely together, with decades of research experience from plasticity and thermo-visco-plasticity, numerical methods, artificial intelligence, and HPC,” said Koric.

Koric, Technical Assistant Director of NCSA and Research Associate Professor in MechSE, led the project. Diab Abueidda, a recent doctoral graduate from MechSE and current post-doctoral affiliate of NCSA, was the project’s alpha and omega. MechSE Professor Huseyin Sehitoglu and Nahil Sobh, an affiliate professor of the Center for Artificial Intelligence Innovation, also provided valuable contributions.