Observation-Model Fusion¶

Why It Matters¶

Humans tend to pick sides - either trust the model or trust the data - when both contain useful information. Ignoring observations means propagating model biases; ignoring models means losing physical consistency and predictive power. Fusion extracts maximum information from both sources, using observations to correct model drift and models to interpolate sparse observations.

The Key Move¶

For any quantity you want to estimate, ask: what does the model predict? What uncertainty does that prediction have? What observations are available? What uncertainty do they have? Use the observations to update the model prediction, weighting by inverse uncertainty (trust precise sources more). This is Bayesian updating: prior (model) + likelihood (observations) = posterior (fused estimate). Repeat continuously as new observations stream in.

Classic Application¶

Data assimilation in numerical weather prediction. Models predict the atmospheric state 6 hours ahead. New observations (satellite, radar, surface stations) arrive constantly but are sparse and noisy. Data assimilation algorithms (Kalman filters, variational methods) fuse observations with model background, weighting by error covariances. The fused analysis initializes the next forecast. Without fusion, models drift; with fusion, forecasts stay anchored to reality while maintaining physical consistency.

Surprising Application¶

Personal skill development. You have a self-model (I'm good at X, weak at Y) based on past performance - complete but potentially biased. You get sparse, noisy observations (feedback from projects, test scores) - real but incomplete. Fusion: weight recent observations heavily (low uncertainty from recency) to update your self-model, while using your model to interpolate in areas without direct feedback. This prevents over-updating from a single data point (noise) while avoiding ignoring evidence because it conflicts with your prior (bias).

Failure Modes¶

Overconfident sources - if you underestimate model or observation uncertainty, fusion gives wrong weights. Correlated errors - if model and observation errors aren't independent (both use the same flawed assumption), fusion amplifies rather than corrects. Bias blindness - fusion corrects random errors well but struggles with systematic biases unless explicitly modeled. Computational intractability - optimal fusion in high dimensions is often impossible; practical methods make approximations that introduce new errors.

Go Deeper¶

Kalnay, Atmospheric Modeling, Data Assimilation and Predictability, Chapters 5-6; Jazwinski, Stochastic Processes and Filtering Theory (1970).