About Me

I am a PhD candidate in the Department of Geography at the University of Florida, Climate Sciences concentration. My research focuses on understanding, detecting, and predicting heat waves and extreme weather events through the integration of numerical climate models with advanced machine learning techniques.

Research Focus

My work addresses critical challenges in climate science by developing data-driven approaches that enhance our capacity to forecast extreme weather events and support climate adaptation strategies. I am particularly interested in:

Heat wave dynamics and thermodynamic drivers
Extreme weather event mechanisms and predictability
Climate variability and change across temporal scales
Earth system feedbacks and teleconnections
Integration of earth system models with machine learning methods

Selected Ongoing Research

A graph neural network stacked on a physical ensemble produces calibrated week 3-4 heat forecasts of opportunity over the United States

Manuscript in preparation, 2026

Heat is the deadliest weather hazard in the United States, and the decisions that blunt its impact are taken two to four weeks ahead, at a lead time where forecast skill is famously scarce. This work asks whether a machine-learned weather model can add decision-relevant value at this range when it is used not as a stand-alone predictor, but as one input to a calibrated probabilistic system.

We build HeatCast, a GraphCast-style mesh graph neural network that issues week 3-4 (W34) distributional forecasts of daily maximum temperature over the contiguous United States. HeatCast probabilities, forecast margins, and predicted spread are combined with a calibrated physical ensemble through a transparent cross-fit logistic stacker.

Verified on 22 years of held-out hindcasts (2002-2023, 950 paired initializations, and 4.6 x 10^8 grid-cell samples), the stacked forecast raises the Brier skill score for W34 heat exceedance by 0.021 above the calibrated ensemble and the area under the ROC curve by 0.041, while remaining sharply calibrated. The improvement is state dependent and concentrates in physically expected windows of opportunity, including La Nina conditions, active MJO phases, anomalously dry soils, and the model's own high-confidence, low-spread forecasts.

Animated HeatCast model output comparing observed and hindcast W34 continuous z-score fields over CONUS — **HeatCast model output.** Animated W34 continuous z-score fields compare observed heat-risk structure with HeatCast hindcasts across CONUS using 14-day averaged lead windows.

Spatial Brier Skill Score maps for HeatCast and HeatCast plus ENS over CONUS — **HeatCast forecast skill.** Spatial Brier Skill Score maps show where machine-learning heat-risk forecasts and ML+ENS stacking add value over climatology and ECMWF ENS baselines.

CONUS case-study maps comparing observed exceedance, stack probability, ENS probability, and anomaly fields — **Inspectable heat-risk fields.** Case-study maps compare observed heat exceedance, stacked forecast probabilities, ENS probabilities, and anomalies across CONUS.

Spatial Clustering of Heatwave Regimes Through ConvAE Latent Space Analysis and Teleconnection Linkages

In review, Geophysical Research Letters, 2026

This manuscript develops an unsupervised deep-learning framework for identifying spatially coherent heatwave regimes across the contiguous United States. Daily maximum-temperature fields from PRISM are transformed into Heat Severity and Coverage Index (HSCI) fields for May through September heatwave days, and a convolutional autoencoder (ConvAE) learns compact latent representations of the spatial heatwave structure.

The ConvAE latent space is projected and clustered to recover data-driven heatwave regimes without imposing fixed administrative or climatological boundaries. The resulting maps show coherent western, central, southern, northern, and eastern heatwave footprints, including transitional regions that are not well represented by fixed regional divisions.

The manuscript also connects the spatial regimes to large-scale teleconnection variability. The strongest regime separation is associated with mid-latitude and high-latitude circulation patterns such as PNA, NAO, PDO, and AO, while summer ENSO provides weaker discrimination across the CONUS heatwave regimes.

Normalized spatial cluster means for ConvAE-derived heatwave regimes across CONUS — **ConvAE heatwave regimes.** Unsupervised latent-space clustering identifies spatially coherent CONUS heatwave regimes from high-resolution HSCI fields.