Sandupal Dutta

PhD Candidate, Earth and Planetary Sciences, Johns Hopkins University. Commander, Indian Navy (Retd).

Portrait of Sandupal Dutta
Earth and Planetary Sciences
Johns Hopkins University
Baltimore, Maryland, USA
sdutta12@jh.edu

On a dark night at sea, I watched the water around my ship turn neon blue-green as the propellers stirred up a colony of bioluminescent algae. That sight stayed with me through two decades at sea, and the tiny organisms behind it are what I study today. I am a PhD candidate in the Morton K. Blaustein Department of Earth and Planetary Sciences at Johns Hopkins University, advised by Prof. Anand Gnanadesikan with co-advisors Prof. Ján Drgoňa and Prof. Somdatta Goswami. My research uses scientific machine learning (SciML) to learn how phytoplankton respond to their environment from 27 years of multi-satellite ocean-colour observations. I used a Fourier Neural Operator (FNO) emulator, weak-form sparse equation discovery and a physics-informed inverse model for my research to improve the ocean biogeochemical models. I expect to defend my doctorate in early 2027.

Before the PhD, I was a meteorological and oceanographic officer in the Indian Navy. I started in an era when weather charts were still drawn by hand, and I watched numerical modelling and then machine learning reshape forecasting. In my final years of service I led an operational forecasting centre that ran weather, ocean and wave models, and introduced data assimilation that measurably improved forecast skill. The work also reached beyond forecasting, into search and rescue, disaster relief and the protection of coastal and marine ecosystems.

Machine learning drew me in, partly because it rekindled a love of mathematics. Serving at remote postings with poor connectivity, I taught myself Python and data science. I went on to build forecasting models used in daily operations, including a hybrid numerical-model and ML approach for thunderstorms and fog, a CNN-LSTM for heavy rainfall, and deep-learning models for wave height. The success of those hybrid methods is why I took a mid-career break to pursue a PhD where machine learning meets ocean science. In 2026, my team DarwinsLabs was named a Pitch Event Runner-Up in the NASA Space to Soil Challenge for SoilSentinel, a CubeSat concept that distils the NASA and IBM Prithvi-EO-2.0 geospatial foundation model into a 1.19M-parameter model small enough to run on board.

Available from early 2027 for research and applied roles in Earth system prediction, ocean and climate science, and operational meteorology.

Research interests: Scientific machine learning · Ocean biogeochemistry · Phytoplankton dynamics · Ocean-colour remote sensing · Equation discovery · Neural operators · Earth system modelling · Numerical weather prediction · Data assimilation · Operational oceanography · Land degradation · Geospatial foundation models

news

SoilSentinel (team DarwinsLabs) named a Pitch Event Runner-Up in the NASA Space to Soil Challenge.
First-author paper on the ecological controls of phytoplankton size classes published in Global Biogeochemical Cycles.
Workshop paper and poster at NeurIPS 2025, Tackling Climate Change with Machine Learning.
Attended the CESM Tutorial 2025 at the NSF National Center for Atmospheric Research (NCAR), Mesa Lab, Boulder, Colorado.
Participants of the Community Earth System Model (CESM) Tutorial, NCAR Mesa Lab, Boulder, 7 to 11 July 2025.
Participants of the Community Earth System Model (CESM) Tutorial, NCAR Mesa Lab, Boulder, 7 to 11 July 2025. Photo: NSF NCAR.
Poster at the Gordon Research Conference on machine learning for actionable climate science.
Participated in the Princeton University and NOAA GFDL node of the WCRP Global km-Scale Hackathon, which focused on GFDL's X-SHiELD kilometre-scale simulations.
Participants of the Princeton University and GFDL Global km-Scale Hackathon, Forrestal Campus, Princeton, 12 to 16 May 2025.
Participants of the Princeton University and GFDL Global km-Scale Hackathon, Forrestal Campus, Princeton, 12 to 16 May 2025.

selected publications

  1. GBC

    Using machine learning to uncover ecological mechanisms controlling abundance of phytoplankton size classes from large-scale observations

    Dutta, S., & Gnanadesikan, A.

    Global Biogeochemical Cycles, 40(4), e2025GB009036, 2026

    Abs

    Phytoplankton size classes (PSCs) and Phytoplankton functional types (PFTs) determine many fundamental biogeochemical processes including nutrient uptake, energy transfer through marine food webs, ocean carbon export, and gas exchange with the atmosphere. Discerning the causes of spatio‐temporal variability of PSCs is a scientific priority for understanding the ocean's role in and response to climate change. This study intends to decipher the relationships between the abundance of PSCs and environmental predictors using explainable machine learning (XAI) techniques. The target variables were PSCs obtained using three different satellite products: size‐resolved phytoplankton carbon from Kostadinov, Milutinović, et al. (2016), https://doi.org/10.5194/os‐12‐561‐2016 , chlorophyll from MODIS divided according to the algorithm of Hirata et al. (2011), https://doi.org/10.5194/bg‐8‐311‐2011 , and a third product from the Copernicus Marine Service. The environmental predictors were nutrients, light, mixed layer depth, salinity, sea surface temperature (sst), and upwelling. The ML algorithm used was the Random Forest Regressor (RFR). XAI techniques were used to discern the relationship between predictors and PSCs abundance. About 85%–95% of the variability of the size classes in the observational data sets was accounted for by environmental variables known to influence phytoplankton biomass. Although different size classes responded similarly to the environmental drivers (with the exception of Copernicus picoplankton) their scale of response varied. The dominant predictors were found to be shortwave radiation, ammonia, dissolved iron and sea surface temperature. The different satellite products show sensitivity to iron, shortwave radiation, sst and ammonia across the same range of values, but with different magnitudes. Copernicus picoplankton is the only product which is positively related to sst.

    Bib
    @article{Dutta2026using,
  author  = {Dutta, S. and Gnanadesikan, A.},
  title   = {Using machine learning to uncover ecological mechanisms controlling abundance of phytoplankton size classes from large-scale observations},
  journal = {Global Biogeochemical Cycles},
  year    = {2026},
  volume  = {40},
  number  = {4},
  pages   = {e2025GB009036},
  doi     = {10.1029/2025GB009036}
}
    DOI
  2. GBC

    Machine learning methods suggest that large regional changes in phytoplankton biomass produced by Earth System Models do not reflect realistic responses to changing climate

    Gnanadesikan, A., Liu, J., Dutta, S., Feole, B., McCarthy, F., & Qian, J.

    Global Biogeochemical Cycles, 39(10), e2025GB008761, 2025

    Abs

    Surface phytoplankton biomass (measured in mol C m −3 ) represents a critical parameter within the Earth System that is measured from space and simulated in Earth System Models. Under climate change, the current generation of Earth System Models agrees that low‐latitude biomass will decline and high‐latitude biomass will increase. However, on a regional scale, the magnitude, phenology and spatial pattern of these changes are highly inconsistent across models. We use machine learning to investigate the sources of the divergence and evaluate the realism of the simulations. We train Random Forests driven by environmental drivers to simulate surface phytoplankton biomass under both pre‐industrial control and SSP5‐8.5 scenarios. Outside the Arctic, the bulk of the changes in biomass are driven by rearrangements in the spatiotemporal distribution of environmental predictors. Large regional changes in models, however, are associated either with unrealistically low pre‐industrial levels of macronutrients or unrealistically strong responses to those macronutrients. Within the Arctic, relationships between environmental predictors and biomass change under global warming. While increased light drives increased biomass, the effect is largest in models with a high nutrient bias. Feeding inputs from an ensemble of models to an emulator trained on observations predicts observed biomass better than the ensemble of the models does, highlighting the fact that models do not produce the correct relationships between environmental predictors and biomass. However, this technique does not yield mechanistically consistent predictions of biomass under climate change. Skepticism of large regional changes in surface phytoplankton biomass produced by individual models is warranted.

    Bib
    @article{Gnanadesikan2025machine,
  author  = {Gnanadesikan, A. and Liu, J. and Dutta, S. and Feole, B. and McCarthy, F. and Qian, J.},
  title   = {Machine learning methods suggest that large regional changes in phytoplankton biomass produced by Earth System Models do not reflect realistic responses to changing climate},
  journal = {Global Biogeochemical Cycles},
  year    = {2025},
  volume  = {39},
  number  = {10},
  pages   = {e2025GB008761},
  doi     = {10.1029/2025GB008761}
}
    DOI
  3. Oceanog.

    Tools in harmony: integrating observations and models for improved understanding of a changing ocean

    Ombres, E. H., Benway, H., Bisson, K., Larkin, A., Perotti, E. A., Wright-Fairbanks, E., Crosswell, J., Dutta, S., et al.

    Oceanography, 38(3), 2025

    Bib
    @article{Ombres2025tools,
  author  = {Ombres, E. H. and Benway, H. and Bisson, K. and Larkin, A. and Perotti, E. A. and Wright-Fairbanks, E. and Crosswell, J. and Dutta, S. and others},
  title   = {Tools in harmony: integrating observations and models for improved understanding of a changing ocean},
  journal = {Oceanography},
  year    = {2025},
  volume  = {38},
  number  = {3},
  doi     = {10.5670/oceanog.2025.e302}
}
    DOI