• Solar atmosphere inversion model: ML model for inverse solar physics, trained on terabyte-scale simulation data, deployed in the Daniel K. Inouye Solar Telescope workflow.
  • SAR satellite foundation model (WV-Net): Self-supervised learning approach trained on 12M+ images; enables retrospective analysis of 9 years of global-coverage data; improves classification and regression by up to 40% across climatology, atmospheric science, and ocean monitoring.
  • Dual-energy X-ray foundation model: Domain-informed self-supervised embedding model consolidating four longitudinal NIH datasets. Outperforms clinical-standard WHO hip fracture risk assessment by 10%.
  • High-resolution solar forecasting: Score-based diffusion models for downscaling numerical climate simulations.
  • Belle-II particle identification: Physics-informed model for Kaon/Pion classification with 97% accuracy.
  • Fish pose estimation: Custom deep learning pipeline cutting video analysis time by 95%.
  • Evolution-informed neural networks: Tree-structured graph neural network for environmental DNA, improving over baselines by up to 30%.