Hybrid physics-ML models for environmental systems: harmful algal blooms, ecological regime shifts, and ocean/climate risk
Hi, I'm Kunal Rathore. I work at the intersection of mechanistic environmental models and machine learning, building systems that use domain physics to guide model architecture rather than treating ML as a black box.
- PhD, Environmental Sciences and Artificial Intelligence, Oregon State University (CEOAS, SEA Lab, advised by Dr. James Watson)
- M.Tch., Mathematical Modeling and Simulation and B.E., Automobile Engineering
- Co-founder, Cetocean — ocean intelligence startup delivering high-resolution SST and environmental data products to the blue economy
- Former: Seagate Technology (NL2SQL/LLM platform for financial analytics), Persistent Systems (time-series forecasting for network anomaly detection)
| 🧪 Project | ✍️ Publication / Writeup | 📋 Summary |
|---|---|---|
| Universal Differential Equations for regime-shift detection | Accepted, Journal of the Royal Society Interface (rsif-2025-1010) | UDE-NN recovers a hidden harvesting-pressure parameter from population abundance data and forecasts ecological tipping points, benchmarked against DE-GP, ARIMA, and LSTM on a stochastic Tilman-Holling system and RAM Legacy fisheries data (Northern Cod collapse) |
| HybridHAB_detection | In preparation, target: Elsevier journal | Compares three atmospheric-correction pipelines (ACOLITE, OC-SAC, C2RCC) for Sentinel-3 OLCI harmful algal bloom detection across five taxa in the Gulf of Maine; shows hybrid ocean-colour + ROMS models beat colour-only under spatial cross-validation (median F1 ≈ 0.745) |
| Hybrid cyanoHAB reservoir forecasting | SSRN preprint (lead author, with Buckner & Watson) | Bayesian Model Averaging over watershed (SWAT/HAWQS) and hydrodynamic (CE-QUAL-W2) model outputs for cyanobacterial bloom prediction in three US reservoirs, using WHO risk thresholds |
| CatRisk_HurricaneLoss | In preparation for EarthArXiv | Open-source probabilistic catastrophe risk model for hurricane insured losses, built on NOAA HURDAT2 data with Weibull hazard fitting, AAL/OEP/TVaR outputs, benchmarked conceptually against CLIMADA and Verisk/AIR |
| W2J | — | Porting CE-QUAL-W2 (the USACE hydrodynamic and water quality Fortran model) to Julia for parallel, differentiable calibration via Enzyme.jl |
| HAB forecasting UDE (MVCO) | — | Replaces the biological module of a coupled bio-physical ocean model with a neural UDE, using ROMS forcing and IFCB observations of Alexandrium catenella at Martha's Vineyard Coastal Observatory |
- Programming: Python, Julia, R
- ML/Scientific ML: PyTorch, scikit-learn, SciML/Lux.jl, Zygote/Enzyme.jl (differentiable programming), Universal Differential Equations, Neural ODEs, Fourier Neural Operators, DeepONet
- Classical ML: RandomForest, CatBoost, LightGBM, HistGradientBoosting, Bayesian Model Averaging
- Environmental modeling: CE-QUAL-W2, SWAT/HAWQS/SWAT-CUP, ROMS
- Remote sensing / geospatial: Sentinel-2/3 (ACOLITE, C2RCC, SNAP), Google Earth Engine, xarray, cartopy
- Data & infra: HPC/SLURM, Parquet, zarr, NL2SQL/LLM pipelines
- Rathore, Buckner, Watson (2026). Hybrid process-based/ML modeling for cyanobacterial bloom prediction. SSRN. doi.org/10.2139/ssrn.7311404
- Universal Differential Equations for ecological regime-shift detection and forecasting. Journal of the Royal Society Interface (accepted, rsif-2025-1010)
- Explainable AI for environmental prediction. IJCAI 2024 XAI Workshop
- IEEE ICECCT 2019 (supporting author)
🚀 Let's connect
🔗 Portfolio: kjrathore.github.io 💼 LinkedIn: linkedin.com/in/kjrathore 📧 kj_rathore@outlook.com