1 of 1

P2.4-154

Leveraging Local Scale Deep Autoencoder-based Models to Characterize Early Times of Global Atmospheric Transport

Fernández-Godino, M. G.¹, Lucas, D. D.¹, Chung, W. T.², Glascoe, L. G.¹, and Myers, S. C.¹

¹Lawrence Livermore National Laboratory, ²Stanford University

GLOBAL SCALE MODELS ~ 10⁴ km

https://lexica.art/aperture

De Meutter & Hoffman, 2020

Fernández-Godino et al. 2022

initial conditions

Wind speed direction fixed

Release location

variable

Two completely different spatial patterns after 90 minutes of continue release

LOCAL SCALE MODEL ~ 10¹km

Deep convolutional neural network-based autoencoders are used to exploit relationships in wind-driven spatial patterns (DATASET 1).
The encoder compresses the dimensionality to 0.02% of the original size, reducing the data dimension.
The full predictive model demonstrates an error of 8%, a figure of merit in space of 94%, and a precision-recall area under the curve of 0.93.

Chung et al., 2022

Fernández-Godino et al., 2023

In this talk, we explore the latest developments in data science and machine learning for atmospheric and geosciences. We begin by discussing the advantages of using local models to predict global behavior. We see that small perturbations in source location can result in very different plumes, and local models provide detailed information on regional gas release dispersion.

We discuss the use of deep autoencoder-based algorithms to learn local scale physics by observing images. Instead of training models using physical quantities, we use images, or pixel information. We present a three-dimensional temporal Unet CNN Autoencoder-based model. The models capture plume location accurately, but we are working in improving it performance on predicting magnitudes. We identify next steps for improvement, including finding the optimal time step, using better machine learning architectures, optimizing the model size and super-resolution.