Chaoyu Zhang

Using Stable Diffusion to Generate New Fusion Devices

Research Assistant, UWPlasma Group, University of Wisconsin, Madison

June 2024 - Present

Supervised by Dr. Rogerio Jorge

• Used CycleGAN, Diffusion and regression models to transform suboptimal stellarator configurations into optimized ones
• Integrated DESC and QSC solvers to validate the configurations by computing physical outputs like iota, magnetic field variance, and gradients of magnetic pressure
• Reduced magnetic field variance and improved axis elongation by applying a custom loss function that minimizes the difference between DESC and QSC output
• Replaced the QSC framework with a ResNet model to rapidly predict near-axis solutions, achieving an 85% accuracy rate

Data Scientist Intern

Hasmo Consulting Company, Data Analytics Department

June 2023

• Designed a MySQL database schema to track user interactions and website analytics
• Appraised Google Analytics 4's functions and algorithms for analyzing user behavior on websites
• Gathered and analyzed websites with similar search terms to the target website for increasing its visibility and ranking with Google Trends and Uber Suggests

Introduction to Modeling and Data Science in Biology

Instructed by Prof. Otto X. Cordero of MIT

May 2021 - Sep. 2021

• Modeled population dynamics using ODEs to study the relationship between population, cell size, nutrients, and ribosomes, and detect the steady state of half-life
• Independent Research: Research on British Epidemic Forecast Based on SIR model
• Established a SIR model and determined the optimal parameters for the model to fit the real situation
• Gathered data on confirmed cases and vaccination rates from government websites with R
• Predicted the epidemic trends and proposed suggestions for adjusting restrictions on COVID-19

Path Smoothing for Robotics Using Optimization Techniques

Nov. 2024

• Developed a hybrid A* pathfinding and Quadratically Constrained Quadratic Program (QCQP) optimization model to smooth navigation paths on game maps, improving robotics natural movement
• Implemented smoothness and proximity loss functions in Julia using Ipopt, ensuring the path remains anchored to the original start and end points while maintaining obstacle avoidance
• Experimented with the model and demonstrated the optimization effect; conducted sensitivity analysis on model parameters to assess trade-offs between path smoothness and proximity

Raindrop Removal Algorithm Based on Diffusion Models and KAN

May 2024

• Implemented a Kolmogorov-Arnold Network (KAN) to extract global and local features, addressing overlapping rain streaks
• Enhanced robustness by incorporating a diffusion model for iterative denoising, simulating raindrop interference and refining images via reverse diffusion
• Achieved superior results in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM)

Comparative Analysis of CNN Architectures with SENet Enhancement

Mar. 2024

• Led a comparative analysis of CNN models (LeNet and ResNet) enhanced with SENet across different resolutions to optimize image classification
• Designed and executed 120 parallel training and testing jobs on MedMNIST using PyTorch and CHTC
• Achieved marked improvements in classification accuracy with SENet-enhanced models; ResNet's accuracy increased from 85% to 92% at minimal computational cost