Section 1.1: The Role of Deep Learning

Deep learning significantly accelerates this complex undertaking.

The Weather Research and Forecasting (WRF) Model

In the U.S., the primary framework for weather predictions is a sophisticated mesoscale model known as the Weather Research and Forecasting (WRF) model. This model caters to a diverse array of meteorological applications, spanning from local to regional scales. As environmental scientist Jiali Wang from the U.S. Department of Energy’s Argonne National Laboratory aptly states, “It captures everything visible outside your window—from clouds to solar radiation, snow, and even how skyscrapers influence wind patterns.”

The WRF model was developed in the late 1990s through collaboration among various governmental agencies and universities, including the National Center for Atmospheric Research (NCAR) and the National Oceanic and Atmospheric Administration (NOAA). Today, it boasts a vast global user base of over 30,000 registered users across more than 150 countries, supported by workshops and tutorials throughout the year. It serves not only as a research tool but also for real-time forecasting, equipped with two dynamical cores and a data assimilation system.

Section 1.2: Navigating Complex Variables

Given its comprehensive nature, the WRF model must navigate countless interrelated weather variables. Unfortunately, encapsulating these intricate relationships through a single set of analytical equations proves challenging. Instead, scientists resort to a method known as parameterization, modeling relationships at a scale larger than the actual phenomena. While this approach can yield adequate predictive accuracy, it also adds considerable computational complexity, making scalability a significant hurdle.

Can deep learning offer a solution to this dilemma? Researchers from Argonne National Laboratory are optimistic. They are working to integrate deep neural networks (DNNs) to replace some of the parameterizations in the WRF model, with the goal of drastically reducing simulation times without sacrificing accuracy.

The first video explores how data-driven techniques are enhancing weather and climate forecasting, presented by Will Chapman from NSF-NCAR.

Subsection 1.2.1: Deep Learning's Impact on Physical Models

Utilizing two decades of model-generated data from WRF, the Argonne team trained their neural networks and subsequently evaluated their ability to replace physics-based parameterizations. If successful, this could lead to the creation of a dense grid—envision it as a three-dimensional representation of Earth and its atmosphere—allowing for precise calculations of variables like wind speed and humidity. The denser the grid, the better the visualization and the greater the potential for localized accuracy in predictions.

The focus of the research is primarily on the Planetary Boundary Layer (PBL), the lowest atmospheric layer most influenced by human activities. This layer, which extends just a few hundred meters above the Earth’s surface, is crucial for understanding broader atmospheric dynamics, including wind velocity, temperature, and humidity.

The Argonne team has fed the DNNs two streams of data totaling 10,000 data points from Kansas and Alaska, spanning eight days. Their findings suggest that DNNs, trained on the relationships between input and output variables, can effectively simulate wind speeds, temperatures, and humidity over time. Remarkably, a DNN trained in one location was able to predict conditions in nearby areas with over 90% correlation, indicating that data from one region can enhance predictions in adjacent regions.

Section 1.3: The Data-Driven Future

DNNs thrive on vast amounts of data, and fortunately, the Argonne Leadership Computing Facility (ALCF) offers an abundance of it. As part of the Department of Energy's user facilities, ALCF utilizes powerful supercomputers to generate meteorological data from intricate simulations. Coupled with high-performance computing resources from the National Energy Research Scientific Computing Center, researchers have access to nearly a petabyte of data, representing 300 years of North American weather history for training their DNNs.