← Back to glossary

Data Assimilation DA

Method Also:Assimilation / Observation Fusion Updated 2026-11-04
Definition

Fusing real observations with the state of a numerical model into the best estimate of the current real world. It corrects the model's own bias and simultaneously provides physically constrained supervision for data-driven models, making it a core component of weather, ocean and structural monitoring systems.

What it is

Data assimilation (DA) fuses observations with model state. The numerical model provides a prior estimate of the state, observations provide the true value at that time (satellites, radar, in-situ sensors, gauges), and the assimilation algorithm combines them into a state estimate more reliable than either source alone - the "initial condition" or analysis field.

That analysis becomes the starting point for the forecast. Assimilation therefore sets the initial error of the forecast and is the first link in the forecast error chain.

Key points

  • Two mainstream algorithm families: variational methods (3D-Var / 4D-Var) perform an

overall least-squares optimization over a time window, using all observations inside it; filters (Kalman and its extensions) update step by step with lower cost and better real-time behavior. Ensemble Kalman filters (EnKF) approximate the covariance with an ensemble, bridging the two.

  • Covariance is the key quantity: observation weights and the relative trust between model

and observation are ultimately determined by error covariances. A badly set covariance (too large or too small) causes over- or under-correction.

  • The observation operator bridges model and observation: it maps model state into

observation space. Uncorrected biases in that operator (model error, observation error) push the analysis systematically toward the model.

  • Assimilation is a natural teacher for data-driven models: analysis fields, or the residual

between model and observation, are physically consistent by construction and are a cheap source of constrained supervision for purely data-driven models.

Common pitfalls / boundaries

  • Assimilation is not noise filtering. It is optimal estimation given known error statistics;

if those statistics are wrong, the analysis is wrong too.

  • More observations do not automatically mean a better result. With uneven observation coverage

(only parts of the domain observed), updates in unobserved areas rely entirely on the prior and the covariance structure, and can smooth away errors where smoothing is unwarranted.

  • "Error dropped after assimilation" must be verified against independent observations.

Evaluating on the same observations used for assimilation gives an optimistic conclusion.

  • Assimilation does not substitute for improving the model. When the prior model has a systematic

bias, assimilation only corrects it within a limited window; long integrations still drift back toward the wrong attractor.

Related terms

Related resources