Introduction to Numerical Weather Prediction

Fundamentals of Numerical Weather Prediction What is Numerical Weather Prediction and Why Does it Matter? Numerical weather prediction (NWP) is the process of using computers to forecast future atmospheric conditions by solving the physical equations that govern how air moves, heats, and exchanges moisture. Rather than relying purely on patterns and experience, NWP captures the physics of the atmosphere in mathematical form and advances these equations forward in time. The basic idea is straightforward: observe the current state of the atmosphere (temperature, pressure, wind, humidity), use physical laws to calculate how these variables will change, and repeat this calculation many times to project several days into the future. This approach works because the atmosphere follows deterministic physical laws—at least over short timescales. The Physical Foundation: Governing Equations The foundation of all NWP models rests on the Navier-Stokes equations, which describe how fluids flow. In the atmosphere, these equations tell us how wind patterns develop and change based on pressure differences, the Coriolis effect, and friction. However, the atmosphere is not just a fluid in motion. Heat and water vapor play critical roles in weather. Therefore, NWP models couple the Navier-Stokes equations with thermodynamic equations that describe: How temperature changes when air rises or sinks How water vapor condenses into clouds and releases latent heat How radiation from the sun and Earth's surface affects temperature Together, these equations form a complete mathematical description of the atmosphere. But there is a problem: these equations are continuous—they describe what happens at every infinitesimally small point in the atmosphere. Computers cannot handle continuous mathematics directly. Discretization: From Continuous to Discrete To make the problem solvable on a computer, we must convert these continuous equations into a discrete form by placing them on a grid. Imagine dividing the entire atmosphere (or a region of it) into a three-dimensional checkerboard of boxes. Each grid point represents the average atmospheric state within one small box. The continuous partial differential equations are then rewritten in terms of differences between neighboring grid points. The grid has two main components: Horizontal grid: Divides the Earth's surface into latitude-longitude cells Vertical grid: Divides the atmosphere into layers, often using pressure levels (upper atmosphere) or height levels (lower atmosphere) For example, a global model might use a horizontal grid spacing of 25–50 kilometers and 50–100 vertical levels. The choice of grid spacing is crucial: finer grids capture smaller weather features like thunderstorms, but they require vastly more computing power. A regional model focused on a smaller area can afford finer resolution because it has fewer grid points overall. Advancing Time: The Time-Stepping Process Once the equations are on a grid, the model advances them forward in small time increments called time steps. In each step (typically seconds to a few minutes), the model calculates how temperature, wind, and moisture change at every grid point based on the values at neighboring points and physical equations. The model then repeats: it uses the updated values to calculate the next time step, and so on. After thousands of time steps, the model has projected the atmosphere hours or days into the future. Why such small time steps? The atmosphere changes rapidly, and numerical methods become unstable if the time step is too large. Specifically, the time step must satisfy the Courant-Friedrichs-Lewy (CFL) condition, which ensures that information cannot "jump over" a grid point in a single time step. For a given grid spacing, smaller grid spacing requires smaller time steps. This is one reason why high-resolution models are computationally expensive—they need both more grid points and smaller time steps. Resolution Trade-offs: Fine Detail Versus Computational Cost One of the central decisions in designing a weather model is choosing the spatial resolution—how finely to divide the grid. A coarse grid (e.g., 100 km spacing) captures large-scale features like cyclones and ridges but misses local details like individual thunderstorms. A fine grid (e.g., 1 km spacing) can resolve convective storms and terrain-forced flows but requires 10,000 times more grid points, dramatically increasing computation time. <extrainfo> Different models make different choices depending on their purpose. Global models typically use coarser resolution (25–50 km) because they must cover the entire planet. Regional models can use finer resolution (1–5 km) because they cover only a limited area. High-resolution models used for specific applications (like storm prediction) use even finer grids, sometimes 100 meters or less. </extrainfo> Data Assimilation: Creating the Initial Condition A perfect forecast requires a perfect initial condition—an accurate snapshot of the entire atmosphere at the starting time. In reality, observations are imperfect and sparse, particularly over oceans and remote regions. Data assimilation solves this problem by blending all available observations with a previous forecast to create the best possible estimate of the current atmospheric state. Sources of Observational Data Weather models draw data from multiple sources: Surface weather stations record temperature, pressure, wind, and humidity at ground level over land Satellites measure radiances (radiation) and provide temperature and humidity profiles over oceans and remote areas where few surface stations exist Radar systems detect precipitation and winds with high spatial resolution over some regions Radiosondes (weather balloons) measure vertical profiles of temperature, humidity, and wind, providing crucial information aloft Aircraft observations contribute data from the upper troposphere and lower stratosphere during routine flights Despite this variety, observations remain sparse. A weather balloon might be launched twice daily from 100 locations worldwide, leaving vast gaps. Satellites help fill these gaps but have their own limitations. This sparsity means the initial condition always contains errors. How Data Assimilation Works The data assimilation process takes observations and blends them with a short-term forecast (called the "first guess" or "background") to produce an analysis—a consistent, complete three-dimensional picture of the atmosphere. The analysis becomes the initial condition for the next forecast. The key insight is that you don't simply use observations as-is. Instead, you weight them based on their estimated accuracy and use physical consistency constraints to fill in gaps. An observation over a single location influences not just that grid point, but surrounding grid points as well, because the atmospheric state must obey the governing equations. Errors in Initial Conditions Even with data assimilation, initial condition errors are inevitable: Observations are imperfect (instruments have measurement errors) Observations are sparse (gaps remain over oceans and mountains) Assimilation methods have limitations These small initial errors don't matter much at first—a 1 °C temperature error over the Pacific Ocean won't noticeably affect tomorrow's forecast. But as the forecast extends forward, errors grow. This error growth is one of the fundamental limits on weather prediction skill. Forecast Reliability: How Long Can We Forecast? One of the most important practical questions is: how far into the future can we accurately forecast? Short-Term Forecasts (1–3 days) During the first 1–3 days after initialization, numerical weather prediction is highly reliable. Initial condition errors are small, and the atmosphere's large-scale patterns (high and low pressure systems) evolve predictably. A 3-day forecast of a major storm's location is typically accurate within 50–100 kilometers. Medium-Term Forecasts (3–10 days) Beyond 3 days, forecast skill decreases noticeably. Initial condition errors amplify, and small uncertainties can grow into large differences. A 10-day forecast captures the general trend (warm or cool, wet or dry) but cannot pinpoint individual storm locations. The atmosphere exhibits chaotic dynamics—tiny differences in the initial state lead to completely different outcomes after several days, much like the "butterfly effect." Long-Term Forecasts (Beyond 10 days) Forecasts beyond two weeks become essentially useless for specific weather events. The initial condition uncertainty has grown so large that multiple forecasts initialized from slightly different observations diverge dramatically. This is not a failure of the model or observations, but a fundamental property of the atmosphere: it is inherently unpredictable beyond this timescale. Sources of Uncertainty Forecast uncertainty arises from multiple sources: Initial condition errors (sparse, imperfect observations) Model approximations (the equations are simplified; small-scale processes are parameterized) Unresolved processes (features smaller than the grid cannot be explicitly calculated) Model Types: Global, Regional, and High-Resolution Different applications require different models: Global Models Global models simulate the atmosphere worldwide using a grid covering the entire planet. They provide the foundation for operational forecasting and supply boundary conditions for regional models. Global models typically use coarser resolution (25–50 km) because covering the entire planet requires enormous computing resources. Examples include the GFS (Global Forecast System) used by the U.S. National Weather Service and the ECMWF (European Centre for Medium-Range Weather Forecasts) model. Regional Models Regional models focus on a specific geographic area (e.g., North America or Europe) and use finer resolution than global models (5–15 km typical). They are "nested" within global models, meaning the global model provides boundary conditions at the edges of the regional domain. This allows them to resolve smaller features while remaining computationally feasible. Higher resolution means regional models can capture details like terrain effects and lake breezes that global models miss. High-Resolution Models High-resolution models use very fine grids (1 km or finer) to explicitly resolve convection (thunderstorms) and other small-scale phenomena. These models are computationally expensive and are typically run only for limited regions and shorter forecast periods. They are valuable for specialized applications like severe weather prediction. <extrainfo> Model Intercomparison Different forecast centers run different models with different physics, resolution, and data assimilation methods. Comparing outputs from multiple models helps identify systematic biases and increases confidence in forecasts. If all models agree on a storm's location, confidence is high. If models disagree, the forecast is more uncertain. </extrainfo> The Forecast Cycle: Continuous Improvement Operational weather prediction runs in a repeating cycle: Initialization: The most recent observations are assimilated to create an initial condition Integration: The model advances forward, generating a forecast (typically 7–14 days ahead) Verification: The forecast is compared to observations as they become available, revealing where the model performed well or poorly Refinement: Feedback from verification guides improvements in model physics, parameterizations, and data assimilation This cycle repeats every 6 or 12 hours (depending on the model) as new observations accumulate and computing power allows. Each new forecast cycle benefits from the latest observations and the latest version of the model.