3 min readData and AI

Hybrid weather twin: blending physics models with machine learning

In May 2022 TempoQuest described a Hybrid Weather Twin: physics-based weather models blended with machine learning. What each kind of model is good at, and where blending is in use today.

The Earth seen from orbit, split in two: the left half drawn as glowing green and blue model data with white contour lines, the right half as real satellite imagery of cloud over land and sea.

First published May 11, 2022 · Updated November 16, 2026

A head of data science who needs weather forecasts now has two kinds of model to choose from. Physics-based models solve the equations of the atmosphere and take a great deal of computing. Machine learning models produce a forecast quickly once they are trained, but they know only what their training data taught them. Most teams end up needing both. In May 2022, TempoQuest described a concept it called the Hybrid Weather Twin, which blends physics-based models with machine learning. This article revisits the idea: what was proposed, what each kind of model is good at, and where blending is in real use today.

What the 2022 concept proposed

A digital twin is a digital replica of a physical system, kept up to date with measurements from the real thing. The 2022 article applied the term to weather. Its hybrid twin paired a physics-based forecast model with data-driven models trained on observations.

The concept placed machine learning at three points in the forecast chain:

  1. Before the model, to select and prepare observations.
  2. Inside the model, to stand in for the most expensive physics calculations.
  3. After the model, to find patterns and outliers in the output and to correct local bias.

It also used the word hybrid for the infrastructure: CPU and GPU, cloud and on-premises, public and private data. The diagram below is from that article.

Layer diagram of the 2022 Hybrid Weather Twin concept. From the bottom up: hybrid infrastructure (CPU and GPU), hybrid deployment (cloud and on-premise), data sources, data transformation and assimilation, hybrid models that pair a WRF physics-based model with machine learning models, post-processing, and API, visualization and report outputs.

The Hybrid Weather Twin was a concept, not a product.

What each kind of model is good at

Physics-based models start from the laws of motion and thermodynamics. Their strengths:

  • They can simulate a situation that has never been observed, because they do not depend on past examples.
  • Their output is complete and physically consistent: every variable in every grid cell.
  • Their behavior can be traced back to a cause, which matters when a forecast is challenged.

Their cost is computing, and fine grids need a lot of it.

Machine learning models learn patterns from historical data. Their strengths:

  • Once trained, they produce a forecast quickly.
  • They can learn the local quirks of a site that a physics model misses, such as a steady bias at one mast.

Their limit is the training data. A model trained on coarse or short records cannot produce detail or extremes it has never seen.

That is why the two approaches complement each other. The physics model can generate the detailed, consistent data that the learning model needs. The learning model can correct the physics model against observations.

Where blending is in use today

Here are three current examples.

Physics-model output as AI training data

Fine-grid detail is not proof of forecast skill. See how to score a forecast, including a bounded historical comparison with lidar observations.

The lesson for any hybrid system is in that pair of findings. The blend beat the reference where the fine-grid run alone did not, and only scoring against observations showed it.

Verification as the referee

Every blend needs a referee. Standard verification scores a forecast against stations, lidar or radar, using measures such as bias, mean absolute error and root-mean-square error. Separately, TempoQuest is developing a Forecast Verification Tool, designed to improve the verification of weather forecasts.

Questions to settle before you build one

  • What will you score against? Choose the observations and the error measure first. Without them you cannot tell whether a blend helps.
  • Is the training data fine enough? The resolution and cadence of the training set cap what the learned model can show.
  • What happens outside the training range? Plan for the record event. Keep a physics-based forecast in the loop for conditions the learned model has not seen.
  • Who runs the physics model? You can run WRF yourself on GPUs with AceCAST, or have the data delivered to your specification.

What to do next

Request custom data, or see AceCAST.

Your forecast

Have a forecast question of your own?

Tell us the decision it has to serve. One short form reaches the team, and we review every request.