SDS Lab PhD Student Wins Google TPU Research Cloud Award for AI Weather Prediction
Diyali Goswami, MS’24, PhD, civil and environmental engineering, who develops machine learning and AI methods for weather prediction at the Northeastern University Sustainability and Data Sciences Laboratory (SDS Lab), has won a prestigious award through a highly competitive international proposal process. Goswami received credits from Google for Tensor Processing Units (TPUs)—custom-designed chips for accelerated AI and machine learning—along with technical support from Google staff. The award comes as weather prediction has become a leading domain for AI, given its significant societal impact and the value of AI-driven and hybrid physics-AI forecasting systems.
Her proposal, “JumpCast: Physics-Guided Precipitation Nowcasting for Extremes,” was selected for support through Google’s 2026 TPU Research Cloud program. Led by Goswami under the supervision of Professor Auroop Ganguly at Northeastern’s SDS Lab and AI for Climate and Sustainability (AI4CaS) initiative, the project was recognized for its scientific and societal relevance at the intersection of AI, weather, and climate extremes. It will receive up to $50,000 in Google Cloud credits and access to TPU infrastructure for large-scale AI training. Google staff expressed enthusiasm for the work and interest in continued engagement as the project develops.
JumpCast develops physics-guided generative AI methods that integrate radar observations, high-cadence satellite imagery, and numerical weather prediction data to improve forecasting of rare, high-impact precipitation events. The framework explicitly models convective initiation, rapid storm intensification, and heavy-tailed precipitation behavior—phenomena that are routinely underrepresented in conventional forecasting systems. By grounding generative AI architectures in physically consistent atmospheric dynamics, the project advances both predictive skill and scientific interpretability for hazardous weather. Applications include flash flood forecasting, emergency response, reservoir operations, and climate resilience planning.
Goswami leads the technical development of the JumpCast framework, including multimodal forecasting architectures and scalable TPU-based training pipelines. The project will release reproducible training infrastructure, evaluation tools, and pre-trained models to support the broader AI-for-weather research community.