Identifying drivers and dynamics of phytoplankton in the Black Sea: application of a neural emulator

Philip A. H. Smith, Anshul Chauhan, Marilaure Grégoire, Luc Vandenbulcke, Filipe Rodrigues, A. Nørlund Christensen, Michael A. St John, Carsten KrogFrontiers in Marine Science 2026 · DOI

TL;DR

This paper introduces a novel deep learning emulator (a U-Net model) that can forecast phytoplankton and nutrient levels in the Black Sea with remarkable accuracy and speed, significantly outperforming traditional methods. Beyond prediction, the model's interpretability reveals key environmental drivers and their varying influence across seasons and regions, offering a powerful tool for ecosystem management and early warning systems.

What Problem Does This Paper Address?

Understanding and predicting the dynamics of marine ecosystems, particularly phytoplankton blooms, is crucial for managing fisheries, detecting pollution, and assessing the impacts of climate change. Traditional process-based ocean models are powerful but computationally intensive, making rapid scenario testing and forecasting challenging. This paper tackles the need for a faster, yet accurate, method to simulate and understand these complex dynamics, specifically focusing on the Black Sea, a region known for its unique environmental challenges like eutrophication and hypoxia.

Methods

The core of this research is a spatiotemporal convolutional U-Net emulator. This deep learning architecture is designed to process 3D data (2D spatial grids over time) and learn complex, non-linear relationships. The U-Net's encoder-decoder structure with skip connections allows it to capture both broad patterns and fine-scale details in the data.

Key methodological innovations include:

  • 3D U-Net Architecture: Specifically adapted to handle time-series of spatial data, enabling the model to learn temporal dependencies alongside spatial ones.
  • Emulator Approach: Instead of solving complex differential equations, the U-Net learns to emulate the behavior of a high-resolution process-based model, offering massive speedups for predictions.
  • Interpretability Tools: The authors employed a suite of sensitivity analysis techniques:
  • Sobol Indices: To quantify the overall contribution of each input variable to the model's output variance.
  • Gradient-based Global Sensitivity Measures (DGSM): To assess the instantaneous sensitivity of the model to small changes in input variables.
  • Local Elasticity: To understand the spatial scale of influence of different drivers.

Datasets

The study utilized simulation data from the NEMO-BAMHBI coupled ocean-biogeochemical model for the Black Sea, covering the period from 1950 to 2014. This dataset served as the "ground truth" for training and evaluating the neural network emulator. No new observational datasets were introduced for training, but the emulator's performance is benchmarked against traditional methods and climatology.

Key Insights and Ideas

  • Deep Learning as a Powerful Emulator: The U-Net emulator demonstrates that deep learning can effectively learn and replicate the complex dynamics of sophisticated process-based models, offering a significant computational advantage.
  • Shifting Driver Importance: The study reveals a fascinating transition in the dominant drivers of phytoplankton dynamics. In the short term, chlorophyll's own past state (autocorrelation) and light availability (PAR) are most influential. However, as the forecast horizon extends, nutrient availability (nitrate and ammonium) become increasingly critical.
  • Spatial and Seasonal Heterogeneity: Driver importance is not uniform. The research highlights significant regional and seasonal variations, with coastal areas and specific seasons showing different dominant influences (e.g., nutrient-driven blooms in the northwest during spring vs. light-driven blooms in summer).
  • Local vs. Broad Influence: Chlorophyll's influence tends to be local and contracts over time, while nutrient influences become more spatially extensive and act over longer lead times, reflecting the transport and diffusion of nutrients.

Key Results

  • High Predictive Skill: The emulator achieved substantial improvements over baseline methods, with a 41% improvement for chlorophyll and 59% for phosphate in terms of RMSE.
  • Speedup: Once trained, the emulator can produce a 16-day forecast in a fraction of a second on a single GPU, a massive speedup compared to the hours or days required by the original process-based model on multiple cores.
  • Spatial Pattern Preservation: The model not only reduces error magnitudes but also effectively captures coherent spatial patterns and regional dynamics, as evidenced by improvements in MS-SSIM and Pearson correlation metrics.
  • Driver Identification: Sensitivity analyses identified chlorophyll, PAR, nitrate, and ammonium as key drivers, with their relative importance varying significantly by region and forecast lead time.

Limitations and Future Work

  • Data Domain Constraints: The emulator's knowledge is limited by the historical simulation data (1950-2014). It may struggle with extreme events or climate change scenarios that fall outside this historical range.
  • Resolution Limitations: The emulator's spatial resolution is inherited from the NEMO-BAMHBI model (~15 km), which may smooth out or underrepresent fine-scale coastal processes and localized hotspots.
  • Vertical Averaging: The study averaged the upper 50m, potentially obscuring important vertical dynamics like the deep chlorophyll maximum.
  • Future Work Suggestions:
    • Incorporate future climate scenario data into training.
  • Integrate physics-informed constraints (e.g., mass balance) into the neural network.
  • Explore hybrid training approaches combining model data with real-world observations.
  • Investigate modeling the full water column to capture vertical dynamics.
  • Explicitly analyze the role of spatio-temporal context predictors.

Summary updated

Abstract

We present a spatiotemporal convolutional U-Net emulator model to forecast phytoplankton chlorophyll concentrations and key nutrient fields (nitrate and ammonium) in the Black Sea, using simulation data from 1950–2014. The emulator achieved substantially higher predictive skill compared to baseline approaches, with a 41% improvement for chlorophyll and 59% for phosphate, while accurately capturing both spatial and temporal variability in biogeochemical dynamics. In addition to forecasting, interpretability of the model was obtained through Sobol sensitivity analysis, complemented by derivative-based global sensitivity measures (DGSM) and elasticity analysis. These revealed pronounced spatial and seasonal variations in the dominant environmental drivers across the basin, enabling exploration of “what-if” scenarios through targeted perturbations of key physical and biogeochemical drivers. Overall, light availability and nutrient concentrations (particularly nitrate, ammonium, and phosphate) emerged as key contributors, with a transition from predominantly light-driven short-term sensitivity toward increasing nutrient influence at longer lead times, modulated by strong regional and seasonal variability. The ability of the model to forecast biogeochemical states and to identify their dominant drivers highlights its potential as an early warning tool for detecting ecosystem changes and supporting adaptive management of the Black Sea.
BibTeX
@inproceedings{smith2026identifying,
  author = {Philip A. H. Smith and Anshul Chauhan and Marilaure Grégoire and Luc Vandenbulcke and Filipe Rodrigues and A. Nørlund Christensen and Michael A. St John and Carsten Krog},
  title = {{Identifying drivers and dynamics of phytoplankton in the Black Sea: application of a neural emulator}},
  booktitle = {Frontiers in Marine Science},
  year = {2026},
  doi = {10.3389/fmars.2026.1760162},
}
Figure