Engineering simulations often require significant computational resources and time, which creates barriers for users and can slow down project timelines. By using machine learning methods, researchers have discovered how to generate accurate, high-resolution simulation results while using significantly less resources.
Researchers from Carnegie Mellon University have developed an up-sampling method, called the Taylor Expansion Error Correction Network (TEECNet). This neural network is effective across a variety of physics problems, including heat transfer and fluid flow. It achieves over 96% accuracy in data enhancement while using 42.76% less computational resources than other popular up-sampling methods. This research is published in the Journal of Computational Physics.
Chris McComb, an associate professor of mechanical engineering, compared TEECNet to the “enhance” button featured in many CSI shows. Much like how that button can improve the resolution of low-quality photos, TEECNet can take data from fast, low-cost simulations and use an algorithm to enhance the quality to that of a more intensive simulation.
TEECNet differs from other up-sampling methods because it prioritizes efficiency.
“We can always learn what we want if models have enough time and data, but we want ours to be efficient and accurate,” said Wenzhou Xu, a Carnegie Mellon University Ph.D. student and lead author of the study.
Noelia Grande GutiΓ©rrez, an assistant professor of mechanical engineering, said they hope to reduce the large data and time costs required by other, less efficient up-sampling methods by embedding more physics knowledge into TEECNet.
TEECNet currently achieves faster results when run on smaller computers. For example, TEECNet-assisted simulations run on computers with 48 cores achieve an average 47.15% cost reduction, while those run on 12-core computers achieve an average 68.77% reduction. Future work will seek to solve this issue to increase the scale of problems that TEECNet can be used to solve.
