Scientific machine learning for ocean biogeochemistry, and operational machine learning from two decades of naval forecasting.
A four-stage SciML pipeline that learns basin-specific governing equations for phytoplankton from 27 years (1998 to 2025) of Copernicus multi-satellite ocean colour and hands them to the E3SM MARBL biogeochemistry model. The North Atlantic spring-bloom equation holds on held-out years (weak R² 0.82), and every coefficient carries bootstrap confidence intervals.
Explainable machine learning (permutation importance, accumulated local effects and SHAP) applied to SeaWiFS, MODIS and OLCI retrievals to separate ecological signal from algorithmic assumptions.
Machine learning shows that the large regional changes in phytoplankton biomass projected by Earth System Models do not reflect realistic responses to a changing climate.
Pitch Event Runner-Up (team DarwinsLabs). The 364.5M-parameter Prithvi-EO-2.0 foundation model was distilled to a 1.19M-parameter model (307 times smaller) for on-board inference, scoring AUC-PR 0.919 over 146 million pixels.
Machine-learning models for anomalous radar propagation, wave-height downscaling, low-level wind shear, crosswind and rainfall, built for and used in day-to-day Indian Navy operations.