Learned 1-D passive scalar advection to accelerate chemical transport modeling: a case study with GEOS-FP horizontal wind fields

TL;DR

This paper introduces a neural network-based 1-D advection scheme that mimics high-resolution solvers, enabling faster chemical transport modeling with high accuracy, and extends to 2-D advection with promising stability and precision.

Contribution

A novel learned advection operator using neural networks that reproduces high-resolution behavior at lower resolutions, improving accuracy over traditional low-res methods.

Findings

High fidelity in reproducing training data

Comparable accuracy on unseen conditions

Achieves 18× acceleration over traditional low-res solver

Abstract

We developed and applied a machine-learned discretization for one-dimensional (1-D) horizontal passive scalar advection, which is an operator component common to all chemical transport models (CTMs). Our learned advection scheme resembles a second-order accuracy, three-stencil numerical solver, but differs from a traditional solver in that coefficients for each equation term are output by a neural network rather than being theoretically-derived constants. We downsampled higher-resolution simulation results -- resulting in up to 16$\times$ larger grid size and 64$\times$ larger timestep -- and trained our neural network-based scheme to match the downsampled integration data. In this way, we created an operator that is low-resolution (in time or space) but can reproduce the behavior of a high-resolution traditional solver. Our model shows high fidelity in reproducing its training dataset (a single 10-day 1-D simulation) and is similarly accurate in simulations with unseen initial conditions, wind fields, and grid spacing. In many cases, our learned solver is more accurate than a low-resolution version of the reference solver, but the low-resolution reference solver achieves greater computational speedup (500$\times$ acceleration) over the high-resolution simulation than the learned solver is able to (18$\times$ acceleration). Surprisingly, our learned 1-D scheme -- when combined with a splitting technique -- can be used to predict 2-D advection, and is in some cases more stable and accurate than the low-resolution reference solver in 2-D. Overall, our results suggest that learned advection operators may offer a higher-accuracy method for accelerating CTM simulations as compared to simply running a traditional integrator at low resolution.