the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 27 Jul 2026
Deep learning model emulators for marine biogeochemistry forecasting from days to decades
Abstract. There has been a major surge of activity in the development of deep-learning emulators for Earth System Models. These emulators offer the potential to substantially reduce computational costs, when used to run short-range, seasonal, and climate-scale predictions. When trained on reanalysis data, they may even outperform conventional numerical model forecasts. Marine biogeochemistry has so far lagged somewhat behind these developments, although a rapid expansion of activity in this area is already happening. We have used a simplified one-dimensional water column model, coupling a physical ocean model with a high complexity marine biogeochemistry model (European Regional Seas Ecosystem Model, ERSEM), and have demonstrated that ERSEM can be successfully emulated using deep learning techniques. We explored two key emulator architectures, Long Short-Term Memory (LSTM) Neural Networks that emulate a selected subset of ERSEM variables at daily time-step resolution, and physics-informed 1D Convolutional Neural Networks (CNN) that emulate the full pelagic ERSEM system throughout the entire water column. We show that by using ocean physics simulator inputs, these emulators remain largely stable over multi-decadal timescales and are highly skilful in reproducing ERSEM simulations in both decadal climate projections and short-range (10-day) forecasting applications. The former includes the emulator's ability to accurately predict the timing of phytoplankton Spring blooms several years in advance. Furthermore, we show that, when trained on reanalysis data, the emulators can outperform ERSEM forecast skill score for several key variables, including phytoplankton and zooplankton, by 50–60 %. If similar performance can be achieved in three-dimensional regional applications, the emulators could deliver substantially higher-quality predictions at a fraction of the computational cost. We also apply novel explainability techniques, which can help provide insights into emulated system's emergent behaviour and feed important information to marine biogeochemistry model builders. The emulator performance is evaluated using a range of metrics, including the ability to reproduce daily anomalies and extreme events. We anticipate that these emulator techniques will have wide applicability in the future, including operational forecasting and in marine autonomous systems. We conclude by discussing key challenges and opportunities for further development.
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-3784', Anonymous Referee #1, 06 Oct 2026
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2026/egusphere-2026-3784/egusphere-2026-3784-RC1-supplement.pdfCitation: https://doi.org/
10.5194/egusphere-2026-3784-RC1 -
RC2: 'Comment on egusphere-2026-3784', Anonymous Referee #2, 15 Oct 2026
The authors apply two machine learning-based emulators to output from a 1-dimensional biogeochemistry ocean model. The results are interesting despite not incorporating 3-dimensional model output. The manuscript is mostly well written and easy to follow.
general comments
The manuscript has numerous figures, with inconsistent placement of units (sometimes missing), inconsistent tick marks, panel labels ("A)"), and legends. Because so many figures are used, some have been placed in the appendix, though some appendix figures are referenced in the text more often than figures included in the main manuscript. Furthermore, some figures could easily be merged and provide additional information to the reader that way. For example, Figs. 3 and 4, as well as Fig. A8B show essentially the same information for 3 different emulator configurations. By merging them together, the reader could more easily compare the performance of the emulators. The panels would become more busy; here I would recommend moving some of the 9 variables/panels into an appendix figure, providing more space for the remaining panels. I would also suggest using a single legend (like in Fig. 2 and not one for each panel as in Fig. 8A), tick marks for each year in every panel (the years do not need to appear under each panel but the tick marks should) and perhaps additional horizontal grid lines (like in Fig. 13) to mark the start of each year.
A second example, which also highlights some of the inconsistencies between figures: Fig. 7 and Fig. A5 can be combined (Fig. A5 falsely states it is the same as Fig. 6, but for the CNN). These two figures also show years on the x-axis, but in a different way than in Fig. 3 and Fig. 4 mentioned above. Here again, it would save space and make the figure more readable by using tick marks but not year numbers for each panel. It would further help to let all panels have the same y-axis to emphasize the difference between variables (currently the reader needs to scan every y-axis to get an idea about the rate of decline of the correlation coefficient). More figures could be merged, and I have more suggested improvements in my specific comments below.
I like that the authors provide an overview of their experiments in Section 2.3, "The experiments". However, the names of the experiments introduced there are too rarely, if ever, used in the manuscript. Later in the text, an experiment is described in much detail again or described without providing much detail at all. In those cases, it would be helpful to refer to Section 2.3 so the reader knows which setup is used.
Similarly, at some point in the manuscript, the authors start using the expression "GOTM-FABM-ERSEM simulator". I like this naming, as it avoids the ambiguous "model" which is sometimes used for the biogeochemistry model (simulator) and sometimes for the neural network/machine learning model (emulator). I would suggest using the simulator and emulator terms more consistently throughout the manuscript after briefly introducing them when first used.
The authors suggest that an issue late in the CNN emulation is triggered by a salinity anomaly in the GOTM-FABM-ERSEM simulation. I would suggest testing this hypothesis by removing the anomaly from the data and replacing it with climatology for that short period.
specific comments (there are no line numbers in the document, so I am using quotes to reference parts of the manuscript)"We have used a simplified" ... "We explored two key emulator architectures": There is a change in the tense here; I would suggest using the simple past when if the sentence describes experiments conducted for the study described in this manuscript, i.e., "We used a simplified..." or perhaps to emphasize "In this study, we used a simplified ...".
"We explored two key emulator architectures, Long Short-Term Memory (LSTM) Neural Networks that emulate a selected subset of ERSEM variables at daily time-step resolution, and physics-informed 1D Convolutional Neural Networks (CNN) that emulate the full pelagic ERSEM system throughout the entire water column.": Here it would be useful to add dimension information for the LSTM approach and perhaps attach the 1D label to the "full pelagic ERSEM system throughout the entire water column" and not the CNN.
"More recently, substantial effort has also been devoted to the development of deep learning emulators": Briefly define what an emulator is; are the AI-based forecasting systems for the atmosphere in the previous sentence emulators as well? Defining emulators would give readers not too familiar with the topic an idea of how AI is employed to perform forecasting.
"emulators trained on reanalysis products": Briefly define or describe these reanalysis products so that readers get the idea why training on observation-informed model output may be better than training on potentially more biased free model output.
"The free run data used to train, validate and test the model were obtained from": This model-generated data is used to train the emulator, correct? I would suggest not using the term "model" here or not by itself, because it could refer to the biogeochemical model or the neural network/deep learning model.
"As shown by Fig.1-Fig.2, this was sufficient.": How does one tell?
Fig 1: Is the black dashed line the climatology? Labels/a legend would be helpful. Tick marks in all panels and a tick for the start of every year (or perhaps a vertical grid line at the stat of each year) would improve the figure. The same comment applies to Fig. A1 and many time-series plots in other figures.
"Similarly to the “RUS” reference scheme from Higgs et al. (2025), the ensemble was used to only estimate the background variances and spatial vertical correlations.": A bit more information would be useful here; how is the observation operator configured, that is, is observed chlorophyll assimilated into a mixed layer average or surface only? It is good to reference a paper that describes the technique in detail, but some basic information can be stated briefly in the manuscript.
"The observational error was set to 15%": I would suggest adding "of the observed chlorophyll value".
"These variables were selected based on both, their importance for the model stakeholders and their relevance for the ecosystem dynamics": The "these" suggests that the variables were mentioned previously, when they were not. As a reader, I was wondering at this point which variables were selected based on their importance; why not list them here? Reading further, I see that they are listed under (c) in the next paragraph where the inputs are listed without mentioning that the (c) inputs are also used as outputs. I would suggest listing the outputs early on. Also mention that they add up to be n_state=20 variables.
"In addition, the forcing variables at the prediction time step are supplied separately [...] The LSTM then predicts the selected biogeochemistry variables (those listed in point c) at the next daily time step.": The two sentences make it sound like the next daily time step is one step ahead of the prediction time step, which I don't think is the case.
"Training samples were created using a sliding window approach.": It is not helpful for the reader to end the previous paragraph with an alternative approach that was not used and then describe the training data generation here. I would suggest mentioning this approach first before going into the alternative training data generation using anomalies.
"State variables, forcing variables, and prediction targets were normalized independently.": Wasn't each variable normalized independently? The sentence makes it sound like the state variables were normalized together but independent of the forcing variables that were normalized together.
"predicts daily averages of 52 ERSEM pelagic variables": I would suggest adding a "the", as in "the 52 ERSEM pelagic variables" to emphasize again that all are being estimated.
"It takes the same daily averaged atmospheric forcing variables as the LSTM network": Mention "input" somewhere.
"It takes the same daily averaged atmospheric forcing variables as the LSTM network [...] and daily averages of temperature and salinity across the full water column": And none of the BGC variables are used as input? This appears to be contradicted by Table 2 listing n_vars as both input and output. This information appears to be missing here.
Table 2 and Table 1: Perhaps in the "Details" or in the "Input Shape" columns, mention the value of n_state, n_forcing, n_vars, d, etc. when first used.
Does the hat over the X imply estimation or normalization? Mention this somewhere.
"Short-range free-run prediction experiments where emulators, trained on" ... "Decadal free-run prediction experiments, where both types of emulators, trained on": Does "emulators" in the "Short-range free-run prediction" imply both types of emulators? Mention in the introductory paragraph that all experiments were performed with both emulators if this is the case.
"meteorological forcing variables": These were previously called "physical forcing variables" or do the two differ?
"the lag at which the mean sensitivity first drops to a given percentage of its peak mean value is identified": This seems to imply that sensitivity values are highest for the shortest lag. This makes sense but could be made more explicit. That is, mention that peak sensitivities are always/typically associated with lags from 0 to n days.
"where the emulator begins to diffuse sharp gradients": It is not fully clear what this means. Do the authors mean the sharp gradients that are present in the simulator output?
"Furthermore, there is considerable interannual variability in the emulated vertical profiles": How does this line up with the previously stated regression towards the mean?
Fig. 6: The x-tick labels are difficult to read.
Fig. 7: It is difficult to compare the panels because each has a different y-axis.
Fig. 7: What does "The Pearson correlation is calculated for the selected biogeochemistry variables." in the caption mean?
Fig. 8: Is the blue line the reference RMSE or the non-normalized emulator RMSE? Here, too, it may be much easier to interpret the figure with shared y-axes for all panels (only the left, relative value).
"Fig.9 shows the skill of the LSTM in forecasting the inter-annual variability": Only with respect to the simulator. I would suggest adding the OC-CCI data to the figure as well. Otherwise, statements such as "These results suggest that emulators driven by ocean physics forecast could be used on a multi-year-to-decadal basis to forecast the onset of algae blooms." rely very much on the simulator doing a good job.
Fig. 9: Is this figure based on the "Seven-year reanalysis forecast" or one of the "free-run prediction" simulations?
"The multiplicative noise was in the most extreme case drawn (independently for each variable) as a white noise": It is useful for the reader to describe the red noise generation, but what does the "in the most extreme case" imply? The procedure was different in less extreme cases?
Fig. 10: Again, consider using the same y-limits for all panels.
Fig. 11: Panels A and B appear to have odd differences in resolution/font size. But why not combine the two panels? Only the emulator results are different; everything else in the panels appears to be the same. Having both emulator results on the same panel should not make the panel much busier and would give it more vertical space. I would suggest moving the emulator results to the front (higher z-order) to give them more emphasis. The same comment applies to Fig. 11 C and D, where, in addition, units are missing from the y-axis.
"The simulator forecast and the two emulator runs from Fig.11 were all started from the reanalysis on the 1st January 2019.": But were they informed by further reanalysis results along the way? I assume not, but it is not entirely clear. I do not understand why the names from the "The experiments" section are not used here.
"Among the predicted outputs, we focus here on three variables of the highest interest: ...": Of highest interest to the authors?
Fig. 13 (caption): "the maximum absolute sensitivity of surface diatom chlorophyll to each input": How is this maximum calculated (I presume it is a maximum across time)? Is the maximum representative of general/average sensitivity?
"In this study, we investigated the feasibility of using deep learning to emulate a highly complex regional marine biogeochemical model": While some of the inputs come from regional model configuration, only a 1-dimensional model is emulated here. And given that there are models that simulate tens or hundreds of plankton variables, I would consider this ERSEM configuration an intermediate complexity model.
Citation: https://doi.org/10.5194/egusphere-2026-3784-RC2
Viewed
Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 163 | 0 | 1 | 164 | 0 | 0 |
- HTML: 163
- PDF: 0
- XML: 1
- Total: 164
- BibTeX: 0
- EndNote: 0
Viewed (geographical distribution)
Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
