A Staged Deep Learning Approach to Spatial Refinement in 3D Temporal Atmospheric Transport

TL;DR

The paper tackles the prohibitive cost of high-fidelity LES for 3D atmospheric plume dispersion in complex terrain by introducing DST3D-UNet-SR, a dual-stage model combining a temporal module with ConvLSTM-enabled 3D U-Net and a spatial refinement module that upscales predictions via 3D U-Net super-resolution. The modular design enables independent optimization of temporal evolution and spatial detail, delivering large speedups (up to three orders of magnitude over LES) while maintaining or improving accuracy compared to a high-resolution temporal baseline. Validation against simulated sensor data shows strong near-field accuracy and robust performance with observational updates, particularly near the source, and favorable mass-conservation properties across time. The work demonstrates a viable path to real-time plume dispersion modeling, iterative optimization, and uncertainty quantification, with future directions including inferring initial fields from wind conditions and incorporating topographic information for better generalization.

Abstract

High-resolution spatiotemporal simulations effectively capture the complexities of atmospheric plume dispersion in complex terrain. However, their high computational cost makes them impractical for applications requiring rapid responses or iterative processes, such as optimization, uncertainty quantification, or inverse modeling. To address this challenge, this work introduces the Dual-Stage Temporal Three-dimensional UNet Super-resolution (DST3D-UNet-SR) model, a highly efficient deep learning model for plume dispersion prediction. DST3D-UNet-SR is composed of two sequential modules: the temporal module (TM), which predicts the transient evolution of a plume in complex terrain from low-resolution temporal data, and the spatial refinement module (SRM), which subsequently enhances the spatial resolution of the TM predictions. We train DST3DUNet- SR using a comprehensive dataset derived from high-resolution large eddy simulations (LES) of plume transport. We propose the DST3D-UNet-SR model to significantly accelerate LES simulations of three-dimensional plume dispersion by three orders of magnitude. Additionally, the model demonstrates the ability to dynamically adapt to evolving conditions through the incorporation of new observational data, substantially improving prediction accuracy in high-concentration regions near the source. Keywords: Atmospheric sciences, Geosciences, Plume transport,3D temporal sequences, Artificial intelligence, CNN, LSTM, Autoencoder, Autoregressive model, U-Net, Super-resolution, Spatial Refinement.

Figures (10)

• Figure 1: An example of a xenon field at 10 m slice above the ground level with $w_s$ = 7.8 m/s, $w_d$ = 357$^\circ$ after 2.3 hours following the start of the release. The white star marks the release location, while the yellow arrow indicates the wind direction. The complex terrain, characterized by mountains and valleys, is shown in gray scale.
• Figure 2: DualStage Temporal 3D UNet-SR Model Architecture
• Figure 3: DST3D-UNet-SR workflow as combination of the TM and SRM. The TM receives a batch of low-resolution sequences and outputs a batch of predicted low-resolution frames. The next TM's prediction input consists of the last four time steps of the previous input concatenated with the latest TM's prediction. The TM's inputs and the predicted low-resolution frames from the fifth element of each sequence are then fed into the SRM, which outputs the corresponding high-resolution frames.
• Figure 4: Averaged concentration in the x-z plane at various time steps, comparing DST3D-UNet-SR and HRTM predictions (two top rows) with ground truth (third row). The errors are also shown in the figure (two bottom rows). The DST3D-UNet-SR error is calculated as the difference between DST3D-UNet-SR predictions and ground truth. Similarly, the HRTM error is calculated as the difference between HRTM predictions and ground truth. Variations are shown across the x and z dimensions within a range of 0 to 5 km and the z-dimension ranging from 0 to 2 km.
• Figure 5: Averaged concentration in the y-z plane at various time steps, comparing DST3D-UNet-SR and HRTM predictions (top two rows) with ground truth (third row). The errors are also shown in the figure (two bottom rows). The DST3D-UNet-SR error is calculated as the difference between DST3D-UNet-SR predictions and ground truth. Similarly, the HRTM error is calculated as the difference between HRTM predictions and ground truth. Variations are shown across the y and z dimensions within a range of 0 to 5 km and the z-dimension ranging from 0 to 2 km.