Multimodal Atmospheric Super-Resolution With Deep Generative Models

TL;DR

The paper demonstrates that a score-based diffusion framework can perform high-dimensional atmospheric super-resolution by fusing multimodal, sparse observations in a zero-shot, Bayesian manner. It presents the EDM-based diffusion model, posterior sampling via diffusion posterior sampling, and a data-fusion approach that accommodates LR ERA5, IGRA radiosonde data, and a learned atmospheric emulator without retraining. Empirical results on ERA5-LR-IGRA-LUCIE data show accurate, spectrally faithful reconstructions and quantified uncertainty, with clear benefits from multimodal inputs and time-consistent trajectories. This approach offers a computationally efficient alternative to traditional data assimilation for rapid, uncertainty-aware state reconstruction in the atmosphere.

Abstract

Score-based diffusion modeling is a generative machine learning algorithm that can be used to sample from complex distributions. They achieve this by learning a score function, i.e., the gradient of the log-probability density of the data, and reversing a noising process using the same. Once trained, score-based diffusion models not only generate new samples but also enable zero-shot conditioning of the generated samples on observed data. This promises a novel paradigm for data and model fusion, wherein the implicitly learned distributions of pretrained score-based diffusion models can be updated given the availability of online data in a Bayesian formulation. In this article, we apply such a concept to the super-resolution of a high-dimensional dynamical system, given the real-time availability of low-resolution and experimentally observed sparse sensor measurements from multimodal data. Additional analysis on how score-based sampling can be used for uncertainty estimates is also provided. Our experiments are performed for a super-resolution task that generates the ERA5 atmospheric dataset given sparse observations from a coarse-grained representation of the same and/or from unstructured experimental observations of the IGRA radiosonde dataset. We demonstrate accurate recovery of the high dimensional state given multiple sources of low-fidelity measurements. We also discover that the generative model can balance the influence of multiple dataset modalities during spatiotemporal reconstructions.

Figures (19)

• Figure 1: Super-resolution using zero-shot posterior sampling given structured grid low-resolution observations (sample-LR). Contours showing mean of samples at specific time instances.
• Figure 2: Comparison of 2m temperature fields from low-resolution input (LR), super-resolution using zero-shot posterior sampling (SR), and bicubic super resolution. The bottom row shows the true field and the corresponding error maps for SR and bicubic methods relative to the true field.
• Figure 3: Comparison of 500mb V-wind fields from low-resolution input (LR), super-resolution using zero-shot posterior sampling (SR), and bicubic interpolation. The bottom row shows the true field and the corresponding error maps for SR and bicubic methods relative to the true field.
• Figure 4: Ensemble based mean (left) and standard deviations (right) obtained from ensembles obtained with sample-LR zero-shot sampling. Showing 2m temperature, 10m u component of wind, 10m v component of wind, and specific humidity at 850mb pressure from top to bottom. One can observe a fine pattern of low uncertainty on a uniform grid, which corresponds to locations for the coarsely sampled ERA5.
• Figure 5: Spectral recovery exhibited by zero-shot sampling from ERA5 generative model when using sample-LR for zero-shot sampling. Note how the proposed approach recovers the right spectral trend as against that of bicubic interpolation