Model-informed ML Estimates European Shelf Carbon Pools

The arXiv preprint 2508.10178, by Jozef Skakala (submitted 13 Aug 2025, revised 17 Jun 2026), describes experiments using a computationally cheap ensemble of neural networks to estimate marine carbon pools in the North-West European Shelf (NWES), per the arXiv abstract. The paper reports training a deep ensemble on a NWES coupled physical-biogeochemistry model free run, then driving the trained ensemble with (a) a NWES reanalysis and (b) the observations assimilated into that reanalysis. The authors report that the ensemble predicts several carbon pools, including detritus, zooplankton, and heterotrophic bacteria, in closer agreement with the reanalysis than the free run, while also producing uncertainty estimates and explainability outputs. The paper notes the ensemble also performs when driven directly by assimilated observations, with the limitation that predictions are then available only at observed locations and times. The arXiv entry records a journal reference: JGR - Machine Learning and Computation 3 (2026).

Per the paper, the core method is a deep ensemble trained to map directly observable inputs, atmospheric, riverine, and ocean variables, to internal carbon-pool fields produced by a coupled model free run. The paper evaluates the ensemble by substituting reanalysis inputs for the free-run inputs and by running on assimilated observation inputs. The authors emphasise explainability and quantify predictive uncertainty from the ensemble spread. The manuscript includes spatial comparisons (example: 2016-2020 average surface concentrations) and short-range forecast assimilation experiments referenced in the abstract and supplementary figures.

Model-informed machine learning, as presented in this work, fits a growing pattern where ML is trained on physics-model outputs to approximate expensive numerical products. For practitioners, this pattern often trades explicit physical simulation cost for a learned surrogate that is cheaper at inference time, while inheriting biases present in the training model. Such surrogates are attractive for operational forecasting pipelines where latency and compute budgets constrain model complexity.

Industry and research observers have explored hybrid ML-data-assimilation workflows for ocean forecasting; the paper situates its contribution within that trend. For marine and climate data teams, a credible, explainable surrogate with uncertainty metrics could reduce reliance on full reanalysis runs for some monitoring or scenario tasks, but it does not remove the need to validate against withheld observations and to quantify how training-model biases propagate into the surrogate.

• •Publication details and peer review outcomes for the JGR article, which will clarify methods and limitations.
• •Independent replication using other shelf regions or alternative biogeochemical models to test generalisability.
• •Work combining deep ensemble surrogates with formal data-assimilation to assess operational forecast improvements.