💻 computer science

Estimating carbon pools in the European Shelf sea environment: replacing reanalysis by model-informed machine learning?

This paper proposes a computationally efficient deep ensemble machine learning approach trained on a physics-biogeochemistry model to accurately estimate European Shelf carbon pools and their uncertainties, offering a viable alternative to expensive reanalyses and a method to complement sparse observations.

Original authors: Jozef Skakala

Published 2026-06-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jozef Skakala

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

The Big Picture: Predicting the Ocean's "Invisible" Carbon

Imagine the ocean as a giant, bustling kitchen. The chefs (phytoplankton) cook up food using sunlight. But the kitchen is messy, and we can only see the chefs through a small window (satellite cameras). We can't easily see the leftovers (detritus), the diners (zooplankton), or the bacteria cleaning up the crumbs. These "invisible" ingredients are crucial for understanding how the ocean absorbs carbon, but they are hard to measure directly.

Scientists usually try to figure out what's in the kitchen by running a massive, complex simulation of the whole restaurant. This is called a Reanalysis. It's like hiring a team of 100 expert chefs to simulate every single meal for a whole year. It's incredibly accurate, but it's also expensive, slow, and requires a supercomputer.

This paper proposes a shortcut: Machine Learning (ML).

The New Approach: The "Smart Apprentice"

Instead of hiring 100 chefs to simulate the whole year, the author trained a "smart apprentice" (a neural network) to learn the rules of the kitchen.

The Training Phase: The apprentice watched a "free run" simulation (a practice session where the computer simulated the ocean without any real-world corrections). The apprentice learned the relationship between what we can see (water temperature, salinity, and the amount of green chlorophyll from satellites) and what we can't see (the carbon pools like bacteria and zooplankton).
The Test: Once trained, the apprentice was given real-world data (the actual satellite observations) and asked to guess the invisible carbon pools.

The Results: Fast, Cheap, and Surprisingly Good

The paper found that this "smart apprentice" is incredibly effective:

Speed: While the supercomputer takes about an hour to simulate just one day of ocean data, the machine learning model can predict years of data in a few seconds on a regular desktop computer. It's the difference between baking a cake from scratch every time versus using a pre-made mix that tastes 90% as good.
Accuracy: When the apprentice was tested against the "gold standard" (the expensive Reanalysis), it did a much better job than the original practice simulation. It successfully predicted the amounts of detritus, zooplankton, and bacteria, matching the expensive model's results much more closely.
Uncertainty: The model didn't just give one answer; it gave a "confidence score." It's like the apprentice saying, "I'm pretty sure there are 50 bacteria here, but I'm only 60% sure about the 51st one."

The "What-If" Scenarios: Simulating the Future

One of the coolest features of this method is that it allows scientists to play "What If?" games very easily.

The Scenario: What if climate change causes the "chefs" (phytoplankton) to disappear? Or what if the type of chefs changes from big ones to tiny ones?
The Experiment: The researchers fed the model a scenario where the phytoplankton slowly vanished. The model predicted how the rest of the kitchen (the carbon pools) would react.
The Catch: The paper notes that if you push the scenario too far (making the phytoplankton disappear completely), the model starts to get a bit confused because it hasn't seen that extreme situation before. It's like asking a chef who only cooks for humans to cook for a ghost; the rules might break down. However, for moderate changes, the model works well.

What It Can't Do Yet

The paper is honest about its limitations:

Depth: The model is great at predicting what's happening at the ocean's surface (the "kitchen counter"), but it struggles to predict what's happening deep down in the "basement" (the deep ocean). It mostly learned the general seasonal patterns of the deep water rather than dynamic changes.
Dissolved Inorganic Carbon (DIC): For one specific type of carbon (DIC), the model didn't improve much over the old simulation. This is likely because the old simulation was already quite good at predicting this specific ingredient.

The Bottom Line

This paper demonstrates that we can use a "model-informed" machine learning tool to act as a fast, cheap, and efficient substitute for expensive ocean simulations. It allows us to fill in the gaps where we don't have direct measurements, helping us understand the ocean's carbon cycle without needing a supercomputer running 24/7. It's a powerful new tool for monitoring the health of our planet's carbon cycle.

Technical Summary: Estimating Carbon Pools in the European Shelf Sea Environment

Problem Statement
Shelf seas, particularly the North-West European Shelf (NWES), play a disproportionate role in the global carbon cycle relative to their size. However, direct observations of organic carbon pools (e.g., detritus, zooplankton, heterotrophic bacteria) in these regions are often sparse, uncertain, or methodologically inconsistent. While marine biogeochemistry models (such as ERSEM coupled with NEMO) can simulate these pools, free-running models often exhibit substantial biases, particularly in phytoplankton phenology, which propagates errors through the ecosystem. Conversely, data assimilation techniques used to create reanalyses (constraining models with satellite and in-situ data) produce more reliable estimates but are computationally expensive, often requiring supercomputing resources and significant wall-clock time. There is a need for a computationally efficient alternative that can reproduce reanalysis-quality carbon pool estimates while retaining the ability to quantify uncertainty and simulate future scenarios.

Methodology
The authors propose a "model-informed machine learning" approach using a deep ensemble of feed-forward neural networks (NNs).

Training Data: The NNs were trained on a 5-year (2016–2020) free-run simulation of the NEMO-FABM-ERSEM coupled physics-biogeochemistry model. The data were coarsened from the operational 7 km/daily resolution to 35 km/10-day resolution to reduce redundancy and improve training efficiency.
Model Architecture: A fully-connected feed-forward network with three hidden layers (1080, 720, and 360 neurons) was implemented using TensorFlow Keras. To estimate epistemic uncertainty and boost performance, a 15-member deep ensemble was employed, utilizing random weight initialization, dropout (30%), and batch splitting.
Inputs: The model ingests observable and structural variables: sea surface temperature (SST), sea surface salinity (SSS), PFT (phytoplankton functional type) chlorophyll-a, short-wave radiation, wind speed, riverine discharge (nutrients, DIC, etc.), and structural data (latitude, longitude, bathymetry, day of year).
Outputs: The model predicts surface and vertically averaged concentrations of five carbon pools: detritus, dissolved organic carbon (DOC), zooplankton, heterotrophic bacteria, and dissolved inorganic carbon (DIC).
Validation Strategy: The trained ensemble was tested against:
1. The independent 2020 portion of the free-run data.
2. The NWES Copernicus multi-decadal reanalysis (which assimilates satellite SST and PFT chlorophyll-a).
3. Direct satellite observations (replacing reanalysis inputs with assimilated satellite data) to simulate the ultimate application.
4. "What-if" scenarios involving reduced phytoplankton biomass and altered community structure.
5. 1D sensitivity experiments (GOTM-ERSEM) with perturbed parameters to assess model structural uncertainty.

Key Results

Performance on Reanalysis: When driven by reanalysis inputs, the deep ensemble successfully reproduced the reanalysis estimates for surface carbon pools (detritus, zooplankton, heterotrophic bacteria, DOC) with significantly higher accuracy than the free-running model. The ensemble captured the reduction in carbon pools caused by data assimilation (which corrects phytoplankton biases), achieving $R^2$ scores between 0.83 and 0.89 for surface pools.
DIC Limitations: The model did not significantly improve upon the free run for Dissolved Inorganic Carbon (DIC). The authors attribute this to the free run already performing well for DIC due to strong constraints from identical drivers (air-sea fluxes, river discharge, SST) in both the free run and reanalysis, leaving little room for ML improvement.
Observation-Driven Prediction: When the model was driven directly by satellite observations (with data gaps aggregated to the required resolution), its skill remained comparable to the reanalysis-driven runs for organic carbon pools, demonstrating its viability as a direct observation-to-pool estimator.
Vertical Profiles: The model failed to improve predictions for vertically averaged pools compared to the free run. SHAP (SHapley Additive exPlanations) analysis revealed that the model learned the seasonal climatology of these variables primarily from structural inputs (coordinates, bathymetry) rather than dynamic inputs, indicating a lack of flexibility in capturing vertical dynamics from surface data alone.
Uncertainty and Scenarios: The deep ensemble provided uncertainty estimates (standard deviation across ensemble members). In "what-if" scenarios (e.g., scaling phytoplankton biomass to zero), the model predicted declines in organic carbon pools but failed to reduce them to zero, indicating it operates outside its training distribution in extreme conditions. 1D experiments confirmed that the model systematically underestimates the impact of severe phytoplankton declines on carbon pools, though it performed reasonably well for bacterial carbon until biomass dropped by >30%.

Significance and Claims
The paper claims that model-informed machine learning offers a viable, computationally efficient alternative to expensive reanalysis systems for estimating surface carbon pools in shelf seas.

Efficiency: The ML model can generate years of predictions in seconds on a desktop, contrasting with the supercomputing requirements of reanalysis.
Gap Filling: It provides estimates for carbon pools (detritus, zooplankton, bacteria) that currently lack robust satellite retrieval algorithms or sufficient in-situ data in the NWES.
Scenario Simulation: The lightweight nature of the model facilitates the rapid simulation of "what-if" climate scenarios, although the authors caution that predictions outside the range of training data (e.g., extreme biomass collapse) may be unreliable.
Modesty: The authors explicitly state that the approach is currently limited to surface pools and coarse resolutions (35 km, 10 days). They acknowledge that the model learns relationships specific to the underlying physics-biogeochemistry model used for training and does not inherently capture the full range of model structural uncertainties, though ensemble methods provide a partial proxy. They conclude that while promising, the transferability to other regions and the ability to predict vertical profiles dynamically remain open questions requiring future investigation.