[Paper Review] Orders-of-magnitude speedup in atmospheric chemistry modeling through neural network-based emulation

Makoto Kelp, Christopher W. Tessum+1 more|arXiv (Cornell University)|Aug 11, 2018

Air Quality Monitoring and Forecasting20 citations

TL;DR

This paper presents a deep neural network emulator for the CBM-Z chemical mechanism in atmospheric chemistry models, achieving a 250x speedup on CPU and 4,250x on GPU with median RMSE below 0.02 ppb. The model accurately replicates one-hour concentration changes across 77 species under diverse conditions, enabling rapid simulation while preserving key chemical dynamics.

ABSTRACT

Chemical transport models (CTMs), which simulate air pollution transport, transformation, and removal, are computationally expensive, largely because of the computational intensity of the chemical mechanisms: systems of coupled differential equations representing atmospheric chemistry. Here we investigate the potential for machine learning to reproduce the behavior of a chemical mechanism, yet with reduced computational expense. We create a 17-layer residual multi-target regression neural network to emulate the Carbon Bond Mechanism Z (CBM-Z) gas-phase chemical mechanism. We train the network to match CBM-Z predictions of changes in concentrations of 77 chemical species after one hour, given a range of chemical and meteorological input conditions, which it is able to do with root-mean-square error (RMSE) of less than 1.97 ppb (median RMSE = 0.02 ppb), while achieving a 250x computational speedup. An additional 17x speedup (total 4250x speedup) is achieved by running the neural network on a graphics-processing unit (GPU). The neural network is able to reproduce the emergent behavior of the chemical system over diurnal cycles using Euler integration, but additional work is needed to constrain the propagation of errors as simulation time progresses.

Motivation & Objective

• To reduce the computational cost of chemical transport models (CTMs), which are limited by expensive chemical mechanism calculations.
• To develop a machine learning surrogate that replicates the behavior of the CBM-Z chemical mechanism with high fidelity.
• To enable rapid simulation of atmospheric chemistry over diurnal cycles while maintaining accuracy.
• To evaluate the feasibility of using deep learning to emulate complex, coupled differential equation systems in air quality modeling.

Proposed method

• A 17-layer residual multi-target regression neural network is trained to predict changes in 77 chemical species' concentrations after one hour of atmospheric processing.
• The network is trained on input data including initial concentrations, emissions, and meteorological variables, with outputs matching CBM-Z model predictions.
• Model training uses a large dataset of CBM-Z simulations across diverse chemical and meteorological conditions.
• The trained neural network is executed on CPU and GPU to assess computational speedup and inference efficiency.
• Euler integration is used to simulate diurnal cycles, testing the emulator's ability to maintain emergent chemical behavior over time.
• Error propagation is monitored to assess long-term stability and accuracy in extended simulations.

Experimental results

Research questions

• RQ1 Can a deep neural network accurately emulate the CBM-Z chemical mechanism across a wide range of atmospheric conditions?
• RQ2 To what extent can a neural network surrogate reduce computational time in atmospheric chemistry modeling?
• RQ3 How well does the emulator preserve the emergent behavior of the chemical system over diurnal cycles?
• RQ4 What is the impact of error accumulation in the neural network emulator over extended simulation periods?

Key findings

• The neural network emulator achieves a median root-mean-square error (RMSE) of 0.02 ppb and a maximum RMSE of less than 1.97 ppb across 77 chemical species.
• The model delivers a 250x speedup when running on a CPU compared to the original CBM-Z model.
• An additional 17x speedup is achieved when the model is run on a GPU, resulting in a total 4,250x speedup.
• The emulator successfully reproduces the emergent behavior of the chemical system over diurnal cycles using Euler integration.
• Error propagation remains a challenge, requiring further work to ensure long-term simulation stability.
• The trained model and training code are available for research use upon request from the corresponding author.