DeepAR: Segmenting Atmospheric Rivers with Segment Anything
Semester project · Jan 2025, Aug 2025, Jan 2026
Built DeepAR, a Segment Anything Model (SAM)-based model to detect atmospheric rivers in climate data. Trained on ERA5 reanalysis, it segments atmospheric rivers well there but struggles on CMIP6 climate projections (SSP2-4.5, SSP5-8.5) due to distribution shift.
A FuXi-Inspired Model for Global Weather Forecasting
Semester project · Jan 2025, Aug 2025, Jan 2026
Built a FuXi-inspired global weather model (Swin Transformer V2 encoder, U-Net decoder) for six-hour-ahead forecasts at 1.5-degree resolution. Compared full fine-tuning against LoRA/PEFT variants; full fine-tuning held up better on long-horizon metrics.
A Local, Low-Resolution Implementation of Pangu-Weather
Semester project · Jan 2025, Aug 2025, Jan 2026
Built a local, lower-resolution (240×121) implementation of Pangu-Weather, Huawei's 3D-transformer weather model, trained on 40 years of ERA5 data. Forecasts stayed usable (ACC above 0.6) for 3–4 days on held-out 2021–2023 test data, despite the resolution cut.
Scaling HEAL-ViT for Medium-Range Weather Prediction
Semester project · Aug 2025, Jan 2026
Built and scaled HEAL-ViT, a HEALPix-based vision transformer for medium-range weather forecasting at 1.5-degree resolution. Scaling the transformer bottleneck from 6 to 48 layers cut RMSE by 12–25% at short-to-medium lead times and improved skill out to 5 days across all 20 forecast variables.
Implementation and Evaluation of GraphCast
Semester project · Aug 2025, Jan 2026
Implemented and evaluated GraphCast, Google DeepMind's graph-neural-network weather model, which learns residual updates on a multi-resolution mesh and rolls them out autoregressively. Trained on 20 atmospheric variables; most maintained usable forecast skill (by ACC) out to about a week.
The Robustness-Accuracy Trade-off in Diffusion-Based Purification
Semester project · Aug 2025, Jan 2026
Derived closed-form expressions for the trade-off between clean and adversarially robust accuracy in a diffusion-based purification framework (DiPure), under a Gaussian mixture model. Found that diffusion noise always degrades clean accuracy, but only improves robustness when the adversarial budget exceeds the signal strength, and even then it can't beat random chance.