Harmful algal blooms can devastate marine ecosystems, contaminate fisheries, and threaten coastal economies. Because these events are driven by complex interactions among ocean currents, weather patterns, and nutrient transport, predicting their occurrence remains a major challenge. In this project, I developed a physics-guided machine learning framework to forecast harmful algal bloom conditions along the California coast. By combining satellite observations, oceanographic data, and physical transport equations, I investigated whether machine learning models could provide accurate forecasts while remaining consistent with the underlying physics governing bloom formation and movement.

Melanoma is responsible for the vast majority of skin cancer deaths, making early and accurate detection critically important. While modern deep learning systems often achieve performance comparable to dermatologists, certain melanomas remain especially difficult to identify because they closely resemble benign lesions. In this project, I investigated how artificial intelligence models classify these challenging “benign-appearing” melanomas and used GradCAM analysis to visualize the image regions influencing model decisions. By comparing model performance on typical and atypical melanoma presentations, I explored potential limitations in current deep learning architectures and identified opportunities for improving future AI-assisted diagnostic systems.

Reliable wind information is critical for sailors, weather forecasting, and marine operations, yet direct measurements are not always available. Inspired by the observation that experienced sailors can often estimate wind speed simply by looking at the water, I investigated whether a neural network could do the same. To answer this question, I assembled a large dataset of ocean-surface photographs paired with wind measurements from NOAA BuoyCAM stations and trained machine learning models to estimate wind speed directly from images. This project explores how computer vision can recover information traditionally inferred by human observation and apply it to real-world environmental forecasting.

Wildfires are becoming increasingly frequent and destructive, creating significant environmental, economic, and public health challenges. While modern machine learning models can predict wildfire risk, they often overlook factors that reduce fire severity, such as firefighting infrastructure and fuel management programs. In this project, I developed deep learning models that incorporate these human inhibition factors into wildfire forecasting across California. The results demonstrate that accounting for human intervention can improve predictive accuracy and provide new insights into strategies for reducing wildfire risk.

Skin cancer affects millions of people each year, yet both physicians and artificial intelligence systems often perform less accurately on patients with darker skin tones because of limited training data. In this project, I developed an artificial intelligence framework that generates synthetic images of skin lesions in darker skin and uses them to train a more equitable diagnostic model. To improve accessibility, I also designed a low-cost handheld device capable of capturing images and providing rapid skin cancer screening, helping expand access to accurate diagnosis across diverse communities.