Hybrid machine learning data assimilation for marine biogeochemistry

arXiv:2504.05218 · physics.ao-ph, cs.LG · Submitted 2025-04-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Hybrid machine learning data assimilation for marine biogeochemistry".

Jane: The paper was written by Ieuan Higgs, Ross Bannister, Jozef Skákala, Alberto Carrassi and Stefano Ciavatta from University of Reading and National Centre for Earth Observation and Plymouth Marine Laboratory and University of Bologna and Mercator Ocean International.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the show, everybody. Today we're digging into a brand new paper from arXiv called "Hybrid machine learning data assimilation for marine biogeochemistry." Jane, I gotta say, that title is a mouthful, but it's got me hooked already.

Jane: It really is a mouthful, Tom, but it's one of those titles that tells you exactly what's inside. We're talking about ocean models, machine learning, and how we blend them together to make better predictions about the health of our seas.

Tom: Right, and when I first saw "marine biogeochemistry," I'll admit, I had to take a breath. But once you break it down, it's basically the study of how living things, chemicals, and the ocean all interact. Think carbon, nitrogen, oxygen, all that stuff.

Jane: Exactly. And this paper is from a team across the UK and Europe, including folks from the University of Reading and Plymouth Marine Laboratory. They're trying to solve a really practical problem: how do we forecast what's happening in the ocean when we can only see a tiny piece of it?

Tom: And that's where the machine learning comes in. It's like having a super-smart assistant that can fill in the blanks. The ocean is huge, we can't measure everything, so we need models to guess what's happening where we aren't looking.

Jane: And that's the real kicker. These models are complex, they're expensive to run, and they're full of variables that we just can't observe directly. So the question becomes, how do we make the best guess possible with the limited data we have?

Tom: So, this paper is essentially saying, "Hey, what if we let AI learn the hidden connections in the ocean, and then use that knowledge to make our forecasts way better?" And that's a pretty exciting idea, if you ask me.

Jane: It is, and it's not just about making prettier models. This has real-world consequences. Better forecasts mean better predictions for fisheries, for harmful algal blooms, for how the ocean absorbs carbon. This is about understanding the planet we live on.

Tom: I love that. So we've got the big picture. Next, we need to get into the nitty-gritty of what they actually did. Stick around, because we're about to break down the abstract and see how they pulled this off.

Abstract: Tom: Alright, we're back with "Hybrid machine learning data assimilation for marine biogeochemistry," and Jane, we just read the abstract, and it's packed. Let's unpack it for our listeners.

Jane: Let's do it. So the paper starts by reminding us that these ocean models are critical for forecasting and understanding climate change. But they have a problem. They're complex, they're non-linear, and the observations we feed them are sparse and uncertain.

Tom: Sparse is the key word there. We're talking about a few satellites measuring chlorophyll at the surface, and maybe some floating sensors, but that's a drop in the bucket compared to the whole ocean.

Jane: Right. And the old-school way of doing data assimilation, which is the process of blending model predictions with real observations, it struggles. It's either too expensive, because you need to run a huge ensemble of models, or it's too simple, and it only updates the variables you directly observe.

Tom: So, the paper's big idea is to use machine learning to learn the statistical relationships between the things we can see, like chlorophyll, and the things we can't, like nitrate or zooplankton. That way, when we get a new observation, the model can make smart updates across the whole system.

Jane: And they tested this in a 1D water column model, which is like a single vertical slice of the ocean. They used it as a testbed to see if the ML could learn those connections. And the results were pretty impressive.

Tom: They were. They showed that their ML approach could significantly improve the updates for variables that were previously just left alone. It's like the model suddenly has eyes on parts of the ocean it was blind to before.

Jane: And they also found that the ML models could transfer to a new location. Not perfectly, but well enough to suggest that you could train one model and use it across a larger area, which is crucial for scaling this up to a full three dee operational system.

Tom: So the abstract is basically saying, "We've found a way to make ocean forecasting better, faster, and more affordable." And that's a big deal. But how did they actually do it? We need to get into the methods.

Jane: Exactly. The abstract gives us the "what," but we need to dig into the "how." And that's coming up next, so stay with us.

Improvements: Tom: Welcome back to the show. We're still on "Hybrid machine learning data assimilation for marine biogeochemistry," and Jane, we've talked about the problem and the big idea. Now let's talk about the specific improvements they're proposing.

Jane: Right. So the paper doesn't just have one idea. It has a couple of different ways to use machine learning. The first is what they call ML-OI, which stands for machine learning optimal interpolation. That's a fancy way of saying the ML learns the correlations between variables.

Tom: So instead of using a static, climatological correlation that's the same every year, the ML model looks at the current state of the ocean and predicts, "Right now, chlorophyll and nitrate are probably linked like this."

Jane: Exactly. It's flow-dependent. It adapts to the current conditions. And then they have a second approach, ML-EtE, which is end-to-end. In that one, the ML model doesn't just learn the correlations; it learns the entire update step.

Tom: So it's like the ML is watching a full data assimilation run and learning, "When this happens, the model should make this exact adjustment." It's emulating the whole process.

Jane: And that's a really clever idea because it bypasses the need to calculate all the error covariances and all that heavy math at run time. The ML just knows what the answer should be.

Tom: And the improvements they saw were significant. For nitrate, which is a key nutrient, they got an eight to twelve percent reduction in error. That's not nothing. That's a real improvement in the model's ability to predict what's happening.

Jane: And they didn't stop at nitrate. They extended it to update almost all the pelagic variables in the model. That's a huge step up from the old univariate schemes that only updated chlorophyll.

Tom: But it wasn't all smooth sailing. They found that zooplankton were tricky. The ML models struggled to update them effectively. It's a reminder that these models aren't magic; they need good training data, and some variables are just harder to predict than others.

Jane: That's a really important point. The paper is honest about the limitations. But the overall message is clear: these hybrid ML-DA schemes are a viable path forward for making marine biogeochemistry forecasting much more effective.

Tom: So we've got the methods and the improvements. But how did they actually test this? We need to talk about the experiments and the results. That's up next.

Page 1: Tom: We're back, still talking about "Hybrid machine learning data assimilation for marine biogeochemistry." And Jane, we've been talking about the ideas, but now we need to get into the nitty-gritty of the setup. Let's look at the first page of the paper.

Jane: Right. The first page sets the stage. It introduces the models they used, which are GOTM for the physics and ERSEM for the biology. GOTM is a 1D ocean turbulence model, and ERSEM is a very detailed marine ecosystem model.

Tom: And ERSEM is no joke. It's got phytoplankton split into four functional types, zooplankton, bacteria, detritus, dissolved organic matter, and a bunch of nutrients. It's a lot of moving parts.

Jane: It is. And that complexity is exactly why data assimilation is so hard. You have all these interacting variables, but you're only observing one thing, which in this case is total chlorophyll at the surface.

Tom: So they set up these 1D models at two locations in the English Channel. One is the L4 site, which is a well-studied, productive coastal site. The other is a less productive site they call CWEC. And they use synthetic observations to test their methods.

Jane: Synthetic observations are a great way to test. You run a "truth" model, then you sample observations from it, and then you see how well your data assimilation scheme can recover the truth. It's a controlled experiment.

Tom: And they used a one hundred-member ensemble to generate the training data for the ML models. That's a lot of model runs, but it gives them a rich dataset to learn from.

Jane: It does. And it's important to note that they're not trying to replace the physical model with machine learning. The ML is only there to improve the data assimilation step. The model still does the forecasting.

Tom: So it's a hybrid approach. You keep the physics, you keep the biology, but you use AI to make the data assimilation smarter. That's a really practical way to think about it.

Jane: Absolutely. And it's a much more achievable goal than trying to get an AI to simulate the entire ocean. This is about augmenting what we already have, not replacing it.

Tom: So we've got the setup. We know the models, we know the locations, we know the data. Now we need to see what actually happened when they ran these experiments. That's the exciting part, and it's coming up right after this.

Conclusion: Tom: And we're back for the final stretch on "Hybrid machine learning data assimilation for marine biogeochemistry." Jane, we've covered a lot of ground, so let's bring it all together.

Jane: Let's do it. The core finding is that machine learning can be a game-changer for marine biogeochemistry data assimilation. It allows us to update unobserved variables effectively, which was a major bottleneck before.

Tom: And they showed it works in two different ways. One where the ML learns the correlations, and one where it learns the entire update step. Both showed real improvements over the standard univariate approach.

Jane: They also showed that the ML models have some transferability. A model trained at one location can be applied to another with reasonable success. That's a big deal for scaling this up to a three dee operational system.

Tom: But they also found that it's not perfect. Zooplankton are still a challenge, and the transfer isn't perfect. It's a reminder that this is a first step, not the final answer.

Jane: Exactly. The paper is very honest about the limitations. But the potential is huge. Better ocean forecasts mean better predictions for fisheries, for harmful algal blooms, for carbon uptake. This has real-world implications for how we manage our oceans.

Tom: And the authors suggest a clear path forward. They talk about using a sparse forest of 1D models to train a system for a three dee domain, which is a clever way to get around the computational cost.

Jane: It is. And it's a testament to the creativity of the researchers. They're not just solving a problem; they're thinking about how to make the solution practical and scalable.

Tom: So, "Hybrid machine learning data assimilation for marine biogeochemistry" is a paper that shows us a smarter way to understand our oceans. It's a blend of physics, biology, and AI that could really change the game.

Jane: And with that, we're going to say goodbye to this paper. It's been a great discussion, and we're excited to see where this research goes next. Thanks for listening, everyone.

Tom: We'll be back with another paper soon. Until then, keep looking up, and keep asking questions. See you next time.

Ieuan Higgs, Ross Bannister, Jozef Skákala, Alberto Carrassi, Stefano Ciavatta

University of Reading · National Centre for Earth Observation · Plymouth Marine Laboratory · University of Bologna · Mercator Ocean International

physics.ao-ph, cs.LG

Submitted: 2025-04-07

Comments: 31 pages, 13 figures (10 in main text, 3 in appendix)

DOI: 10.5194/bg-23-315-2026

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 69/100

Key concepts

Marine Biogeochemistry
This refers to the study of how living things, chemicals, and the ocean interact. It includes tracking elements like carbon, nitrogen, and oxygen within marine ecosystems.
Data Assimilation
This is the process of blending model predictions with real-world observations. The paper discusses how machine learning can improve this process by making smarter updates to model variables based on limited data.
ML-OI (Machine Learning Optimal Interpolation)
This is an ML approach where the machine learning model learns the statistical relationships between observable variables, such as chlorophyll and nitrate. It adapts these learned correlations based on the current state of the ocean.
ML-EtE (Machine Learning End-to-End)
This approach uses ML to learn the entire data assimilation update step directly. Instead of calculating complex error covariances, the ML model learns what adjustment to make when a specific event occurs in the model run.

Terminology

Summary

Summary

This paper investigates the use of machine learning (ML) to improve data assimilation (DA) for marine biogeochemistry (BGC) models, addressing the challenges of model complexity, strong nonlinearity, and sparse, uncertain observations. The authors state: “This study demonstrates how machine learning (ML) can improve marine biogeochemistry DA by learning statistical relationships between observed and unobserved variables.”

The work is set within a 1D prototype of a system used to forecast marine biogeochemistry in the North-West European Shelf seas, coupling the Generalised Ocean Turbulence Model (GOTM) with the European Regional Seas Ecosystem Model (ERSEM) via the Framework for Aquatic Biogeochemical Models (FABM). Two locations are used: L4 (50.25° N, 4.217° W), a highly biologically productive site with seasonally stratified dynamics, and the Central Western English Channel (CWEC) (49.40° N, 4.217° W), which is less productive and less influenced by riverine outflow.

The authors propose two ML-based DA schemes, both building upon a reference univariate scheme (RUS) that mimics operational systems by updating only total chlorophyll (and its constituent phytoplankton functional types via stoichiometric balancing). The first scheme, ML-OI (machine learning optimal interpolation), uses an artificial neural network (ANN) to predict state-dependent correlations between observed total chlorophyll and unobserved variables, which are then combined with climatological variances in the DA update equation. The second scheme, ML-EtE (end-to-end), uses an ANN to directly predict analysis increments for unobserved variables, emulating the updates produced by an Ensemble Kalman Filter (EnKF) training run. Both schemes are compared against the RUS, a purely climatological correlation scheme (CliC), and a stochastic EnKF of varying ensemble sizes.

The key results are as follows. First, for updating a single unobserved variable (nitrate), the ML-OI model vastly outperforms climatological correlation estimates in an offline test, with an RMSE of 0.255 versus 0.731, an approximate 60% improvement. In online cycled experiments, both ML approaches reduce nitrate analysis error by 8–12% compared to the RUS, while the CliC scheme degrades performance. The ML-EtE scheme provides the best forecast improvements, with benefits persisting for approximately 4–5 days of lead time.

Second, when extending updates to all pelagic variables, the ML-EtE (ALL) scheme generally improves unobserved forecast and analysis RMSEs by 10–50% relative to the RUS. The ML-OI (ALL) scheme performs similarly but suffers more from poor zooplankton updates. Excluding zooplankton from updates (ML-OI Excl. Zoo) yields the best overall performance, as zooplankton variables are poorly predicted due to their weak correlations with total chlorophyll.

Third, regarding transferability, the ML-EtE (ALL) scheme performs poorly when applied to the unseen CWEC location, with many RMSEs exceeding 1.5 times the RUS background error. In contrast, the ML-OI (ALL) scheme shows moderate transferability, likely because predicted correlations are more location-agnostic, and when combined with location-specific climatological variances, produce more appropriate increments. The ML-OI (Excl. Zoo) scheme performs best at the new location.

The authors conclude: “ML offers a clear pathway to overcome current computational bottlenecks in marine biogeochemistry DA and that refining transferability, optimizing training data sampling, and evaluating scalability for large-scale marine forecasting, should be future research priorities.” They also discuss a methodology for scaling to 3D operational systems, suggesting the use of a sparse forest of 1D models covering different biogeochemical regimes, and note that “the results suggest that some sparsity could be applied in extracting training data for these models, as long as each regime of BGC behaviour is represented in the selection.”

Improvements for AI systems

Based on the paper, here are the specific improvements I can make to AI systems and what the improved system can do:

  • Improvement: Integrate a two-stage ML pipeline into existing DA systems: (a) an ML model that predicts state-dependent background error correlations between observed and unobserved variables, and (b) an end-to-end ML model that directly predicts analysis increments from the background state and observed increment.

  • What it can do: Update unobserved variables (e.g., nitrate, phosphate, detritus) in real-time without running expensive ensembles, reducing computational cost by orders of magnitude while maintaining or improving forecast accuracy.

  • Improvement: Replace climatological correlations (which are static and smoothed) with a fully connected ANN that learns non-linear, time-varying correlations between total chlorophyll and other pelagic variables, using the current model state as input.

  • What it can do: Capture inter-annual variability in bloom timing and nutrient dynamics, improving the accuracy of multivariate updates by 60% (RMSE reduction from 0.731 to 0.255) compared to climatology.

  • Improvement: Train a neural network to emulate the full EnKF analysis step, taking as input the background state and the total chlorophyll increment, and outputting analysis increments for all unobserved variables.

  • What it can do: Replicate the benefits of a 100-member EnKF at a fraction of the cost, reducing nitrate forecast error by 8–12% and improving forecasts for up to 5 days lead time, without needing to run an ensemble operationally.

  • Improvement: Implement a hybrid scheme (ML-OI Excl. Zoo) that updates all pelagic variables except zooplankton, based on SHAP analysis showing zooplankton have weak correlations and degrade performance when updated.

  • What it can do: Avoid introducing imbalance into the system, achieving the best overall forecast RMSE improvements (10–50% reduction) across all variables, while preventing degradation in zooplankton and downstream ecosystem components.

  • Improvement: Use ML-OI (correlation prediction) rather than ML-EtE (increment prediction) for spatial generalization, as correlations are more location-agnostic. Combine with location-specific climatological variances to generate appropriate increments.

  • What it can do: Apply a model trained at one site (e.g., L4) to a new site (e.g., CWEC) with different biogeochemical behavior, achieving moderate improvements in nitrate, phosphorus, and detritus without retraining, enabling scalable deployment across 3D operational domains.

  • Improvement: Use Kernel SHAP to identify which input features are most important for each prediction target, allowing for pruning of irrelevant features (e.g., oxygen, zooplankton) and better understanding of model decisions.

  • What it can do: Reduce input dimensionality, improve model interpretability, and provide insights into which ecosystem variables are dynamically coupled, guiding future model design and observation strategies.

  • Operational Forecasting: Provide daily 7-day forecasts of marine ecosystem variables (e.g., chlorophyll, nutrients, oxygen) with improved accuracy, especially for unobserved variables, at computational costs suitable for operational centers.

  • Reanalysis Products: Generate multi-decadal reanalyses with multivariate consistency, capturing bloom timing and nutrient dynamics more accurately than current univariate schemes.

  • Scalability to 3D Models: Deploy ML models trained on sparse 1D locations across a 3D domain, using ecoregion-based sampling to cover diverse biogeochemical regimes, enabling basin-scale multivariate DA without prohibitive ensemble costs.

  • Adaptive Observation Assimilation: Respond in real-time to changes in the system state (e.g., storms, fronts, bloom onset) by predicting flow-dependent correlations, rather than relying on static climatology.

  • Uncertainty Quantification: Provide ensemble-like uncertainty estimates for unobserved variables through ML-predicted correlations, without the need for a running ensemble.

  • Transfer Learning: Apply trained models to new locations with minimal retraining, reducing the need for expensive DA training runs at every grid point.

Abstract

Marine biogeochemistry models are critical for forecasting, as well as estimating ecosystem responses to climate change and human activities. Data assimilation (DA) improves these models by aligning them with real-world observations, but marine biogeochemistry DA faces challenges due to model complexity, strong nonlinearity, and sparse, uncertain observations. Existing DA methods applied to marine biogeochemistry struggle to update unobserved variables effectively, while ensemble-based methods are computationally too expensive for high-complexity marine biogeochemistry models. This study demonstrates how machine learning (ML) can improve marine biogeochemistry DA by learning statistical relationships between observed and unobserved variables. We integrate ML-driven balancing schemes into a 1D prototype of a system used to forecast marine biogeochemistry in the North-West European Shelf seas. ML is applied to predict (i) state-dependent correlations from free-run ensembles and (ii), in an ``end-to-end'' fashion, analysis increments from an Ensemble Kalman Filter. Our results show that ML significantly enhances updates for previously not-updated variables when compared to univariate schemes akin to those used operationally. Furthermore, ML models exhibit moderate transferability to new locations, a crucial step toward scaling these methods to 3D operational systems. We conclude that ML offers a clear pathway to overcome current computational bottlenecks in marine biogeochemistry DA and that refining transferability, optimizing training data sampling, and evaluating scalability for large-scale marine forecasting, should be future research priorities.

Related papers