Numerical Weather Prediction (NWP)

The physical core of every NWP model is a set of non-linear partial differential equations governing atmospheric dynamics. These include the Navier-Stokes equations for momentum, the thermodynamic energy equation, the continuity equation for mass conservation, and the ideal gas law. These equations are discretized onto a three-dimensional grid and integrated forward in time using numerical methods like the semi-implicit semi-Lagrangian scheme.

Processes occurring at scales smaller than the model grid resolution cannot be explicitly resolved and must be parameterized. Critical parameterizations include:

• Cumulus convection: Vertical heat and moisture transport by clouds.
• Microphysics: Formation and fallout of precipitation particles.
• Planetary boundary layer: Turbulent mixing of heat and momentum near the surface.
• Radiative transfer: Shortwave and longwave heating rates. These parameterizations are the dominant source of systematic model bias.

An NWP model cannot forecast accurately from an incorrect initial state. Data assimilation fuses a short-term background forecast with millions of heterogeneous observations—radiosondes, satellite radiances, aircraft reports, and surface stations—to produce an optimal analysis. Modern systems use 4D-Var (four-dimensional variational assimilation) or Ensemble Kalman Filters to constrain the analysis to physically consistent dynamics.

Global models (e.g., ECMWF IFS, GFS) cover the entire sphere at moderate resolution (9–13 km). Limited-area models (e.g., HRRR, AROME) operate at convection-permitting scales (1–3 km) over a specific domain. These high-resolution nests receive lateral boundary conditions from a driving global model, creating a one-way nesting dependency. Errors in the global boundary conditions propagate inward and limit the skill horizon of the regional model.

A single deterministic forecast is inherently fragile due to the chaotic nature of the atmosphere. Ensemble prediction runs multiple perturbed forecasts simultaneously to estimate the probability density function of future states. Perturbations target:

• Initial conditions: Using singular vectors or bred vectors.
• Model physics: Stochastically perturbing parameterization tendencies. The ensemble spread quantifies forecast uncertainty, enabling risk-based decision-making for grid operators.

Operational NWP is among the most computationally demanding civilian workloads. A global ensemble system requires sustained petaflop-scale performance on high-performance computing (HPC) clusters with tightly coupled interconnects. The strict time-to-solution constraint—a forecast must complete before the weather it predicts arrives—imposes a hard real-time deadline. This computational wall clock limit constrains the maximum feasible resolution and ensemble size.