Machine Learning Researcher

Built a Bayesian SDE research prototype to reconstruct eelgrass biomass trajectories from sparse data.

What I Did

I developed a non-stationary Bayesian stochastic differential equation model that reconstructs eelgrass biomass time series from only 6 observations per year. The model uses daily satellite data (sea surface temperature and photosynthetically active radiation) as environmental drivers. I implemented joint posterior inference over both the latent biomass states and the process parameters, then validated the uncertainty estimates using CRPS (continuous ranked probability score) and posterior predictive checks.

Impact

The research prototype demonstrated that sparse ecological data can be meaningfully interpolated using physics-informed priors and environmental covariates. The model provided a framework for estimating depletion risks for specific eelgrass sites. The $225B and $110/tonne figures reflect the potential scale of the natural capital assets the broader program aimed to protect.

What I Learned

I gained hands-on experience with stochastic differential equations, specifically implementing Euler-Maruyama discretization and handling non-stationary drift terms. I learned to formulate joint inference problems where latent states and parameters are estimated simultaneously. Working with CRPS for proper scoring rules taught me how to rigorously evaluate probabilistic forecasts beyond simple point accuracy metrics.

Key Highlights

• Prototyped a non-stationary Bayesian SDE model to reconstruct sparse biological trajectories (N=6) from daily satellite covariates (SST, PAR), implementing custom joint posterior inference to rigorously validate forecast reliability via CRPS metrics.

• Leveraged this probabilistic framework to stratify depletion risks for coastal 'Natural Capital' assets ($225B), identifying thermal thresholds to protect blue carbon stocks valued at $110/tonne under 2026 federal carbon pricing schedules.