A NESTED FULLY DATA DRIVEN AFNO-BASED\\ REGIONAL WEATHER FORECAST MODEL

Abstract

In recent years, machine learning (ML) and artificial intelligence (AI) has found applications in a wide variety of fields including geosciences. Driven both by a need for faster and more accurate computation of traditionally slow numerical methods and the increasing availability of large datasets, a variety of AI models and techniques have begun to be applied to the domains of bias correction and forecasting. Poor air quality (AQ) can have a significant impact on human health, which emphasizes the need for accurate air quality forecasts. Traditionally, a Kalman filter is used to correct biases (or errors) in numerical air quality forecasts, but this approach struggles with extreme air quality events such as wildfires and heat waves due to its underlying assumptions. An AI approach can learn how to identify extreme events and then correct them accordingly without relying on these same assumptions. In this work we describe a method that combines unsupervised learning and a forecast-aware bi-directional Long Short-Term Memory (LSTM) network to perform bias correction for operational air quality forecasting using AirNow station data for ozone and PM\textsubscript{2.5} in the continental US. Using an unsupervised clustering method trained on station geographical features such as latitude and longitude, urbanization, and elevation, the learned clusters direct training by partitioning the training data for the LSTM networks. LSTMs are forecast-aware and implemented using a unique way to perform learning forward and backwards in time across forecasting days. When comparing the Root Mean Squared Error (RMSE) of the forecast model to the RMSE of the bias corrected model, the bias corrected model shows significant improvement - 27\% lower RMSE for ozone - over the base forecast. The recent availability of long-term reanalysis datasets such as ECMWF ERA5 and CERRA has enabled the development of AI-driven machine learning models for weather forecasting. AI models can produce weather forecasts significantly faster than traditional numerical methods (seconds or minutes instead of hours or days). Several models have been produced that produce low resolution global forecasts, but higher-resolution regional forecast models have yet to follow. In this work, we present an AI regional forecast at 5.5 km spatial resolution employing the Nvidia FourCastNet (FCN) model with its Adaptive Fourier Neural Operator (AFNO) and transformer self-attention modeling approach. We describe the training of a regional FourCastNet model on the NASA Center for Climate Studies (NCCS) Adapt cluster at the Goddard Space Flight Center using five years of CERRA reanalysis data at 3-hour intervals for five variables across four pressure levels. We show the RMSE forecast errors of a 5.5km implementation trained on 5 years of data are lower than a similar model trained on 5 years of global reanalysis data interpolated to the CERRA domain, demonstrating the ability of our model to learn and forecast regional-scale atmospheric dynamics. Finally, we propose a nesting scheme wherein a model trained on global data is used to produce boundary conditions for a model trained on regional data.