Numerous efforts have been made to accelerate chemical transport models (CTMs) through machine-learning surrogates. While chemistry mechanism reduction has been of particular interest in this field since chemistry modules are the most computationally expensive, transport operators are the second most expensive in many CTMs. Our previous work demonstrated computational speedup of 2-D advection by using spatiotemporal coarse-graining with a convolutional neural network (CNN) solver which emulates the flux terms of a discretized advection equation. The previous approach was incapable of producing stable simulations when temporal coarsening was performed in the native spatial resolution. This was due to spurious noise development in the region where the background concentration is zero. However, solutions that can achieve temporal coarsening without corresponding spatial coarsening would be more useful. In this work, to tackle this noise problem we introduced a local max pooling filter which gives 0 as output if the local maximum concentration is zero or gives 1 as output otherwise. This stencil filter suppressed the noise development, and the new solver achieved a stable 10-day-long 2-D simulation. Temporal coarsening greater than 4× achieved computational speedup against the baseline reference solver with reasonable accuracy (e.g., r2 = 0.79 for 4×). However, temporal coarsening greater than 16× deteriorated the model accuracy (e.g., r2 = 0.35 for 16×). We also employed a Kolmogorov Arnold Network (KAN)-based solver returning stencil coefficients used to compute flux terms. Our preliminary results reveal the KAN solver achieved better accuracy than CNN-based solver in the tested coarsening resolutions (e.g., r2 = 0.82 for 4×). Although the KAN solver achieved better accuracy with less number of parameters, it was still slower than CNN solver. We are working to improve the speed and accuracy of both approaches.
Speaker Bio: Manho Park is a Ph.D. student in the Department of Civil and Environmental Engineering at the University of Illinois Urbana-Champaign, working with Dr. Chris Tessum. He received his bachelor’s and master’s degrees in Civil and Environmental Engineering at Seoul National University in South Korea. His current research interest is to integrate data-driven science and numerical methods to advance transport operators of air quality models.