Geophysical forecasting has profound implications for economic planning, disaster management, and adaptation to climate change. Traditionally, forecasts for geophysical applications have relied on experimental observations, statistical analyses, and high-performance computing. But the computational costs of these methods are high, and equation-based forecasts cannot capture all the relevant physical processes of the atmosphere or the oceans.
Researchers have begun applying machine learning models, including neural networks, to geophysical data to allow accurate, less computationally expensive forecasting. Construction of these neural networks can be challenging and involve significant trial and error.
To address these limitations, Argonne is developing a new tool with the unwieldy name POD-LSTMs (proper-orthogonal decomposition-based, long short-term memory networks) to forecast the global sea-surface temperature using a National Oceanic and Atmospheric Administration (NOAA) data set. The team used DeepHyper, a parallel neural architecture search framework, to obtain better LSTM neural networks through evaluations over multiple compute nodes of a supercomputer. They improved the scalability of the approach using an evolutionary algorithm that improves networks using crossovers and mutations and demonstrated its efficacy on up to 512 nodes of the Theta supercomputer at the Argonne Leadership Computing Facility.
Results from this study yielded neural network forecast models that were as accurate as state-of-the-art equation-based techniques at a fraction of the cost. While the latter requires the solution of partial differential equations on several computers in parallel, the former (once fit) can be executed on a standard single-node laptop.