ML4RT
Machine learning for radiative transfer
We use machine learning to emulate the Rapid Radiative-transfer Model (RRTM), which is commonly used as a radiative transfer (RT) parameterization in numerical weather prediction (NWP). The RRTM itself is an emulator of the full line-by-line RT models, which are the most accurate but are many orders of magnitude too slow for NWP. However, the RRTM is also slow and therefore cannot be called at every time step in NWP. Thus, the RRTM is called only once every Nth (e.g., 20th) time step, with the parent NWP model assuming that RT quantities (fluxes and heating rates) remain constant during the intervening time. This assumption can decrease the overall accuracy of weather predictions. Also, even with the compromise of running only once every Nth time step, the RRTM still accounts for a significant percentage of the computing resources used by the model.
Our solution is to emulate the RRTM with deep neural networks (NN), specifically the U-net++ and U-net3+ architectures, which are specially designed for predicting gridded quantities (e.g., a full vertical profile of heating rates). The NNs are much faster at execution time than the RRTM, so if successfully integrated into NWP, the NN-based emulators could be called every time step. The ultimate goal of this project is to integrate the NNs into the Global Forecast System (GFS) model; we have developed the NNs, and our NOAA collaborators are currently working on the GFS integration. For the most recent iteration of this work (the NNs we plan to integrate into the GFS), see Lagerquist et al. (2023b). For an earlier iteration of this work — emulating only a simplified version of the shortwave RRTM, rather than the full shortwave and longwave — see Lagerquist et al. (2021).
Furthermore, we have used the RT problem as a sandbox for two important theoretical topics in ML: the accuracy/complexity trade-off and uncertainty quantification (UQ). Regarding the first topic, more complex models tend to be more accurate but less interpretable — i.e., it is difficult to understand the physical relationships learned by a complex model or assess its trustworthiness. We are collaborating with Tom Beucler at the University of Lausanne to write a publication on this topic (Beucler et al. 2023b), and we have given conference presentations (Beucler et al. 2023a; Lagerquist et al. 2023a). Regarding the second problem, we are using the RT problem to investigate how UQ methods handle out-of-sample data, e.g., data from a completely different climate or weather regime than the ML model’s training data. It is well known that ML models without UQ struggle to generalize to out-of-sample data, producing poor predictions. However, it is often assumed that UQ will alert users to these poor predictions — i.e., that the poor predictions will be made with low confidence. We show that this is generally not the case, i.e., out-of-sample data lead to poor predictions made with high confidence. In other words, the model silently fails, which could have serious effects in operations. We also explore how to prevent silent failure. This topic is very timely and important, as ML-based UQ is becoming widespread in environmental science. We are currently writing a publication on this topic (Lagerquist et al. 2023c).
Below are two figures from Lagerquist et al. (2023a), showing how different NN architectures — from very simple to very complex — perform in different situations.
Publications
Beucler, T., A. Grundner, R. Lagerquist, and S. Shamekh, 2023a: “Systematically generating hierarchies of machine-learning models, from equation discovery to deep neural networks.” Conference on Artificial Intelligence for Environmental Science, Denver, Colorado, American Meteorological Society, https://ams.confex.com/ams/103ANNUAL/meetingapp.cgi/Paper/414020.
Beucler, T., A. Grundner, S. Shamekh, and R. Lagerquist, 2023b: “Systematically generating hierarchies of machine-learning models, from equation discovery to deep neural networks.” Artificial Intelligence for the Earth Systems, in preparation.
Lagerquist, R., D. Turner, I. Ebert-Uphoff, J. Stewart, and V. Hagerty, 2021: “Using deep learning to emulate and accelerate a radiative-transfer model.”
Journal of Atmospheric and Oceanic Technology, 38 (10), 1673-1696, https://doi.org/10.1175/JTECH-D-21-0007.1.
Lagerquist, R., D. Turner, J. Stewart, I. Ebert-Uphoff, and N. Wang, 2023a: “How complex must neural networks be to accurately estimate radiative transfer in different situations?” Conference on Artificial Intelligence for Environmental Science, Denver, Colorado, American Meteorological Society, https://ams.confex.com/ams/103ANNUAL/meetingapp.cgi/Paper/421334.
Lagerquist, R., D. Turner, I. Ebert-Uphoff, and J. Stewart, 2023b: “Estimating full longwave and shortwave radiative transfer with neural networks of varying complexity.” Journal of Atmospheric and Oceanic Technology, conditionally accepted.
Lagerquist, R., and I. Ebert-Uphoff, 2023c: “Machine-learned uncertainty quantification is not magic: Lessons learned from an application to atmospheric radiation.” Journal of Advances in Modeling Earth Systems, in preparation.
Primary contact
Ryan Lagerquist (CIRA, NOAA GSL)