Machine learning data assimilation brings two complementary approaches together: data assimilation estimates a system’s state from observations, while machine learning identifies patterns that a model may not capture. This combination is relevant to complex, sensor-driven systems, including industrial process models, where measurements may be incomplete or noisy and the underlying model may have persistent error. A quasi-geostrophic model study illustrates one way to build and test this kind of hybrid approach.
Why combine machine learning with data assimilation?
Machine-learning models can approximate the evolution of a system by learning from examples. In a straightforward supervised-learning setup, the model receives a system state at one time and learns to predict the state at a later time. This works best when the training examples accurately represent the system’s behavior.
In many real-world settings, however, the complete system state is not directly available. Measurements may come from only some locations, arrive at intervals, or include noise. Data assimilation addresses this gap by combining observations with a model to estimate the system’s evolving state. In numerical weather prediction, this helps recover a coherent state from imperfect observations; in industrial settings, the same broad challenge arises when process data provide only a partial view of equipment or operations.
Data assimilation can also provide a framework for training a machine-learning model. The estimation problem can be written with two sets of unknowns: the system trajectory and the machine-learning parameters. Optimizing both simultaneously may be difficult for high-dimensional systems, so an alternative is coordinate descent: alternate between estimating the trajectory with data assimilation and updating the learned model with machine learning.
Correct the model rather than replace it
Rather than asking a neural network to learn an entire system from scratch, a hybrid model retains an existing model and adds a trainable correction. In simplified form, the hybrid prediction is the original model’s output plus a machine-learning estimate of its error.
This distinction matters. The learned component is trained to represent the difference between what the existing model predicts and what the system does. The established model therefore remains part of the forecast, while machine learning targets its residual error. For engineers maintaining process models, this is a useful way to think about augmentation: the objective is not necessarily to discard a known model, but to investigate whether a data-driven correction can account for systematic discrepancies.
The approach also makes the quality and meaning of the training target important. If the target is model error, the model learns what the original model missed. If the target is a state increment produced by data assimilation, that increment is not automatically identical to model error: it can also reflect state-estimation effects. The distinction becomes clear in the quasi-geostrophic experiments.
What the quasi-geostrophic experiments tested
The study used a two-layer quasi-geostrophic model, a simplified dynamical system with 1,600 state variables. A perturbed version of the model served as the baseline, while a more accurate version supplied the reference behavior. The perturbations included changes to model parameters and integration time step. The goal was to train a neural network to correct the baseline model’s forecasts.
In an idealized experiment, training data came from complete, noiseless model states. The network learned the error in the perturbed model and was then added to it as a hybrid correction. Across the tested configurations, larger networks generally achieved lower forecast error. Longer training trajectories improved forecasts at an eight-day lead, and shorter sampling intervals generally helped at that lead. The results were not identical at longer lead times: increasing the training period or reducing the sampling interval did not improve the 16-day forecast error.
A key finding was that learning the error of an imperfect model was easier than learning the full system dynamics from persistence alone. The hybrid model based on the perturbed quasi-geostrophic model performed better overall, particularly with smaller training databases. This supports testing corrections in the context of an existing model rather than assuming a standalone machine-learning model is the only option.
Using sparse, noisy observations changes the task
The idealized setup assumes information that is rarely available in operational systems. To test a more realistic case, the researchers generated observations at 50 randomly selected locations, with measurement noise, and applied strong-constraint 4D-Var to estimate the system state over successive windows. The resulting analyses—state estimates produced by data assimilation—were used to create training examples for the neural network.
The network learned about 87% of the variance in the assimilation increments. But when evaluated against the actual model error, it learned about 70% of that variance. This gap shows why an analysis increment should not be treated as a perfect measurement of model error. It is a useful training signal derived from observations and a model, but it may contain more than the correction the machine-learning component is intended to learn.
Even with that limitation, the trained correction improved forecasts compared with the original perturbed model. Its effect was less strong than the correction trained on complete, noiseless observations, as might be expected given the more limited information available. The result demonstrates the central trade-off: data assimilation makes learning possible with sparse, noisy observations, but uncertainty in the estimated state can affect the quality of the learned correction.
Practical considerations for engineering teams
For teams exploring machine learning for process or equipment models, the study suggests several questions to address before deployment:
- Define the learning target. Decide whether the network should approximate the full system behavior, model error, or an assimilation-derived increment.
- Preserve the baseline model. A hybrid structure allows the learned component to focus on correcting an existing model rather than replacing it outright.
- Test data coverage. Training-window length and sampling frequency affected results, and the best choice depended on the forecast horizon.
- Validate beyond training loss. Evaluate the hybrid model in forecast conditions and distinguish performance on training targets from performance against the underlying system behavior.
- Account for integration complexity. Alternating data-assimilation and machine-learning steps avoids one large joint optimization, but convergence is not guaranteed. Applying a hybrid model inside an assimilation cycle can also require model derivatives and coordination between software tools.
Conclusion
Machine learning data assimilation offers a structured way to use observations to improve a model when the full system state is not known. In the quasi-geostrophic tests, a neural network learned useful corrections to a perturbed model, including when training data came from assimilation of sparse, noisy observations. The results also highlight limits: analysis increments are not the same as model error, and improvements depend on the training data and forecast horizon. For industrial automation teams, the practical takeaway is to treat machine learning as a targeted, carefully validated model correction—and to assess the complete hybrid system, not just the network’s training score.
Access Official Technical PDF