Aoyang Li

Experience

1. Research Intern

   National Center For Supercomputing Applications

   Aug 2024 — Present · Urbana, IL, US

   Bidirectional Soil Spectral ML PlatformDesigned end-to-end ML system mapping soil properties to and from hyperspectral reflectance (350–2500 nm).• Forward pipeline: Built multi-task learning pipeline tolerant to missing values, improving R² by 72.9%(sand), 49.6%(phosphorus),42.8%(CEC).• Reverse pipeline: Adapted transformer-based model (MST++) for 1D spectral reconstruction, achieving average R² 0.52 on limiteddata after adding a feature embedding layer to the front.• Integrated LLM-based natural language interface for domain experts to query soil and spectral profiles.Process-Guided Machine Learning for Soil Moisture EstimationCombined radiative transfer modeling with neural nets to deliver daily 9 km global soil moisture product (2015–2025).• Implemented custom physics-guided loss functions, large-scale pre-training (RTM data), and fine-tuning on 360k ground measure-ments. Outperformed seven existing satellite products by 35–50% in RMSE.