Abstract
Modeling atmospheric chemistry is complex and computationally intense. Giventhe recent success of Deep neural networks in digital signal processing, wepropose a Neural Network Emulator for fast chemical concentration modeling. Weconsider atmospheric chemistry as a time-dependent Ordinary DifferentialEquation. To extract the hidden correlations between initial states and futuretime evolution, we propose ChemNNE, an Attention based Neural Network Emulator(NNE) that can model the atmospheric chemistry as a neural ODE process. Toefficiently simulate the chemical changes, we propose the sinusoidal timeembedding to estimate the oscillating tendency over time. More importantly, weuse the Fourier neural operator to model the ODE process for efficientcomputation. We also propose three physical-informed losses to supervise thetraining optimization. To evaluate our model, we propose a large-scale chemicaldataset that can be used for neural network training and evaluation. Theextensive experiments show that our approach achieves state-of-the-artperformance in modeling accuracy and computational speed.