🤖 machine learning

Hybrid machine learning data assimilation for marine biogeochemistry

This study demonstrates that integrating machine learning into data assimilation for marine biogeochemistry significantly improves the updating of unobserved variables and offers a computationally efficient pathway to overcome current limitations in operational forecasting systems.

Original authors: Ieuan Higgs, Ross Bannister, Jozef Skákala, Alberto Carrassi, Stefano Ciavatta

Published 2026-08-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ieuan Higgs, Ross Bannister, Jozef Skákala, Alberto Carrassi, Stefano Ciavatta

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the ocean as a giant, invisible kitchen where tiny chefs (plankton) are constantly cooking up life. To understand what's happening in this kitchen—how much food is being made, how the temperature is changing, or if the pantry is running low on ingredients—scientists use complex computer models. These models are like recipe books that try to predict the future of the ocean's ecosystem. But here's the problem: the ocean is huge, and we can only peek inside it at a few specific spots using satellites or ships. It's like trying to guess the taste of a massive stew by only tasting one spoonful from the top.

To fix this, scientists use a technique called "Data Assimilation." Think of it as a smart sous-chef who constantly tastes the spoonful we can see and then adjusts the entire pot of stew to match that taste. If the spoonful tastes salty, the chef assumes the whole pot needs more salt. However, in the ocean, the ingredients are tricky. We can easily see the "green stuff" (chlorophyll) from space, but we can't easily see the invisible nutrients like nitrate or the tiny animals (zooplankton) eating the green stuff. Traditional methods often struggle to guess how the invisible ingredients should change based on the visible ones, leading to a stew that tastes a bit off. This is where the question of the day comes in: Can we teach a computer to be a better sous-chef, one that understands the secret recipes connecting the visible green stuff to the invisible nutrients?

This paper, titled "Hybrid machine learning data assimilation for marine biogeochemistry," tackles exactly that challenge. The authors, a team of researchers from the UK and Europe, wanted to see if Machine Learning (ML)—the kind of AI that learns from patterns—could help these ocean models make better guesses. Instead of just updating the green stuff they can see, they wanted the AI to learn how to update the invisible ingredients too, like nitrate, phosphate, and different types of plankton.

To test this, the team set up a digital simulation of the ocean in the English Channel. They created a "virtual ocean" that was 100% real in its physics but used made-up data to train their AI. They compared three different ways of adjusting the model:

The Old Way (Univariate): Only updating the green stuff (chlorophyll) and letting the rest of the stew stay as is.
The "Gold Standard" Way (Ensemble Kalman Filter): A very expensive, super-computational method that runs hundreds of simulations at once to guess the connections. It works great but is too slow and costly to run every day for real-world forecasts.
The New AI Way (Hybrid ML): Using a smart computer program to learn the connections between the green stuff and the invisible nutrients, then using that knowledge to update the model quickly.

The researchers tried two specific AI tricks. The first, called ML-OI, taught the AI to predict the relationship (correlation) between the green stuff and the nutrients. For example, "When the green stuff spikes, the nitrate usually drops." The second trick, ML-EtE, was even more direct: it taught the AI to look at the current state of the ocean and the green stuff, and then simply predict the exact amount of change needed for the nutrients, mimicking the expensive "Gold Standard" method without actually running the expensive simulation.

The results were quite promising. When the AI was tested on the location where it was trained (a spot called L4), it did a fantastic job. It learned the complex, non-linear dance between the seasons, the plankton blooms, and the nutrients. In fact, the AI methods improved the accuracy of the nutrient predictions by about 8% to 12% compared to the old method that didn't update nutrients at all. This is a big deal because it means the model can now "see" the invisible ingredients much better, just by looking at the green stuff. The AI was particularly good at catching the timing of spring blooms, a time when the ocean's chemistry changes rapidly.

However, the story isn't a perfect fairy tale. The paper also tested if this AI could work in a different part of the ocean (a spot called CWEC) that it had never seen before. Here, the results were mixed. The "Gold Standard" AI (ML-EtE) struggled to adapt to the new location, essentially because it had memorized the specific "flavor" of the first spot too well. The other AI method (ML-OI), which learned the general relationships rather than specific numbers, did better at transferring its knowledge, but it still had trouble with certain ingredients, specifically the zooplankton (the tiny animals). The AI just couldn't figure out how to update them correctly in either location, suggesting that these tiny creatures are too unpredictable for the current AI to handle.

The authors conclude that while Machine Learning offers a clear path to making ocean forecasts faster and more accurate—solving the problem of being too expensive to run the "Gold Standard" methods every day—it's not a magic wand yet. The AI works well when the ocean behaves similarly to where it was trained, and it needs more work to handle the trickiest parts of the ecosystem, like the tiny animals. The paper suggests that in the future, we could train these AI models on a "forest" of different 1D ocean slices to cover the whole 3D ocean, but for now, this study proves that AI can successfully learn the secret recipes of the ocean, making our digital sous-chefs a whole lot smarter.

Technical Summary: Hybrid Machine Learning Data Assimilation for Marine Biogeochemistry

1. Problem Statement

Marine biogeochemistry (BGC) models are essential for forecasting ecosystem responses to climate change and human activities. However, Data Assimilation (DA) in this domain faces significant hurdles compared to physical oceanography:

Complexity and Nonlinearity: BGC models (e.g., ERSEM) involve tens of variables and hundreds of parameters with highly non-linear interactions.
Sparse and Uncertain Observations: Observations are limited, often restricted to satellite-derived total chlorophyll, while many state variables (nutrients, zooplankton, detritus) remain unobserved.
Computational Constraints: Ensemble-based methods like the Ensemble Kalman Filter (EnKF), which provide flow-dependent error covariances necessary for effective multivariate DA, are computationally prohibitive for high-complexity BGC models in operational settings.
Limitations of Current Schemes: Operational systems often rely on "univariate" schemes that update only observed variables (chlorophyll) and use simple, static balancing principles (e.g., stoichiometry) to update unobserved variables. These approaches often fail to propagate information effectively, leading to marginal improvements or even degradation of unobserved variables.

The core challenge is to develop a DA framework that can effectively update unobserved variables using sparse observations without incurring the computational cost of large ensemble runs.

2. Methodology

The study proposes a hybrid approach integrating Machine Learning (ML) into the DA workflow. The research utilizes a 1D prototype of the North-West European Shelf (NWES) system, coupling the Generalised Ocean Turbulence Model (GOTM) with the ERSEM biogeochemical model.

2.1 Experimental Setup

Locations: Two sites in the English Channel were used: L4 (highly productive, seasonally stratified) for training and Central Western English Channel (CWEC) for testing transferability.
Observations: Synthetic observations of surface total chlorophyll were assimilated weekly.
Baseline Schemes:
- RUS (Reference Univariate Scheme): Updates only total chlorophyll using climatological error variances; other variables are updated via static stoichiometric balancing.
- EnKF: A "gold standard" ensemble-based DA used for comparison and to generate training data.

2.2 Hybrid ML Approaches

Two distinct ML strategies were developed to introduce flow-dependency and multivariate updates:

ML-OI (Machine Learning Optimal Interpolation):
- Mechanism: An Artificial Neural Network (ANN) predicts the state-dependent error correlations ( $COR_{i,c}$ ) between the observed total chlorophyll and unobserved variables (e.g., nitrate) based on the current background state.
- Integration: These predicted correlations are combined with climatological standard deviations to compute analysis increments via the standard DA update equation.
- Training: The ANN is trained on a 100-member ensemble of free-runs to learn the non-linear mapping between state variables and their correlations.
ML-EtE (Machine Learning End-to-End):
- Mechanism: The ANN is trained to directly predict the analysis increments ( $\Delta x$ ) for unobserved variables.
- Inputs: The background state and the analysis increment of the observed variable (total chlorophyll).
- Training: The network learns from the output of an EnKF run, effectively emulating the EnKF's update step in an "end-to-end" fashion.

2.3 Model Architecture

Algorithm: Fully connected ANNs optimized using AutoKeras with Bayesian optimization for hyperparameter selection.
Features: Input features include surface states of pelagic variables (scaled by climatological maxima) and, for ML-EtE, the chlorophyll analysis increment.
Interpretability: SHAP (SHapley Additive exPlanations) values were used to analyze feature importance and understand the dynamical links learned by the models.

3. Key Results

3.1 Single Variable Update (Nitrate)

Correlation Prediction: The ML-OI model significantly outperformed climatological statistics in predicting the time-varying correlation between chlorophyll and nitrate (RMSE reduction from 0.731 to 0.255). It successfully captured complex seasonal regimes (bloom, nutrient-limited, light-limited) and inter-annual variability that static climatology missed.
Analysis and Forecast Skill: In "online" cycled DA, both ML-OI and ML-EtE reduced the analysis error for unobserved nitrate by 8–12% compared to the RUS baseline.
Forecast Persistence: Improved forecasts persisted for 4–5 days. ML-EtE generally outperformed ML-OI, though both showed error growth beyond 5 days, highlighting the difficulty of partially updating highly non-linear systems.

3.2 Multivariate Extension (All Pelagic Variables)

Performance: Extending updates to all pelagic variables (ML-EtE (ALL) and ML-OI (ALL)) generally improved forecast and analysis RMSEs by 10–50% for most variables.
Zooplankton Challenge: Both ML schemes struggled to accurately update zooplankton variables, often resulting in higher errors than the baseline. SHAP analysis revealed that zooplankton states were poor predictors for their own correlations or increments, likely due to weak coupling with chlorophyll and higher trophic-level uncertainty.
Optimization: Excluding zooplankton from the update set (ML-OI (Excl. Zoo)) yielded the best overall performance, suggesting that hybrid schemes should selectively update variables with strong, learnable correlations.

3.3 Transferability to New Locations

Generalization: When models trained at L4 were applied to the CWEC location:
- ML-EtE (ALL) failed significantly, with errors exceeding 1.5x the baseline. The direct emulation of EnKF increments proved too specific to the training location's variability and scale.
- ML-OI (ALL) showed moderate transferability, particularly for detritus and nutrients, as it learns more location-agnostic correlation structures. However, it still struggled with zooplankton.
Implication: While partial transferability exists, direct end-to-end emulation of increments is not yet robust across distinct biogeochemical regimes without retraining or specific adaptation.

4. Significance and Claims

The paper claims that hybrid ML-DA offers a viable pathway to overcome the computational bottlenecks of marine BGC data assimilation.

Efficiency: ML methods can learn complex, flow-dependent statistical relationships from offline ensemble data, enabling multivariate updates in single-model runs without the runtime cost of large ensembles.
Effectiveness: The study demonstrates that ML can significantly improve the propagation of information from observed (chlorophyll) to unobserved (nutrients, detritus) variables, outperforming traditional univariate schemes and static climatological corrections.
Scalability Pathway: While 3D operational implementation requires further work, the authors propose a methodology using a "sparse forest" of 1D models trained across different biogeochemical regimes to generate the necessary statistics for 3D DA systems.
Future Priorities: The authors identify refining transferability strategies, optimizing training data sampling for ergodic coverage, and evaluating scalability for large-scale 3D systems as critical future research directions.

The study concludes that while ML does not replace the dynamical BGC model, it provides a clear mechanism to enhance the DA step, making multivariate assimilation feasible for complex marine systems where ensemble methods are currently too expensive.