Aimee Zeng

Climate downscaling with foundation embeddings

Built a probabilistic model that downscales ERA5 25 km precipitation grid to 1.5 km in four recursive 2× steps, using AlphaEarth satellite embeddings.

• Developed a Conditional Flow Matching model (U-Net with cross-attention) and XGBoost baselines that produce calibrated ensembles rather than a single guess.
• Engineered a pipeline joining ERA5/ERA5-Land, PRISM, and a 400 GB+ embedding store streamed from cloud storage, with multi-phase training on NCSA Delta AI.
• Evaluated with RMSE, CRPS, coverage, and spread-skill against 800 m PRISM observations.

Biodiversity-aware solar siting

Solar farms need a lot of open land, and the cheapest land isn't always the land wildlife can spare. This tool picks sites that balance energy against habitat cost across seven wildlife guilds, formulated as a binary integer program on a 1 km grid. A second phase will add deep learning forecasts of land-use change.

Multisensor remote sensing for cover crop biomass

Cover crops protect Midwest soil over winter, but measuring how much they grew means clipping plots by hand. We estimate biomass from Sentinel-2/HLS optical imagery, Sentinel-1 radar, proximal sensor readings, and PRISM climate data, trained against ground samples from the Great Lakes Cover Crops Project.

Causal effects of phthalate exposures

People are exposed to several phthalates at once, at different levels, and the exposures interact. I work on estimating how those combinations affect middle aged women's menopausal hot flashes from observational data.

• Applied factorial balancing weights to estimate causal main and interaction effects.
• Developed a framework for multi-level factors and extended it to incomplete designs, where some treatment combinations are never observed.
• Validated the framework with simulation and bootstrap inference.

Deforestation and carbon sequestration at Baima Snow Mountain

I did research on Yunnan snub-nosed monkey, an endangered primate, whose habitat was losing in the Baima Snow Mountain Nature Reserve. I interviewed villagers and the reserve's guards, then mapped the reserve in QGIS and found deforested land inside the core zone, the part that's supposed to be the most protected.

• Located three deforested patches in the core zone, 0.674 km² in total, from satellite imagery in QGIS.
• Combined GlobBiomass forest biomass data with carbon storage estimates for cold-temperate coniferous forest to estimate what reforesting those patches could sequester: about 4,800 tons of CO₂.
• Priced that carbon on China's new national carbon market at roughly 239,000 RMB, as a way to fund reforestation and the reserve's underpaid guards.