Lixing Wang

4D Scilicon Tracker Performance Characterization with ML

Research with European Organization for Nuclear Research (CERN) | Advisor: Prof. Gaetano Barone | Jun 2025 - Present

Developed a comprehensive system to characterize Low-Gain Avalanche Detector (LGAD) performance under environmental stress factors using machine learning. Worked with post-docs and grad students on data collection and detector manufacturing.

See IEEE Poster here.

Key Contributions:

• Built a robust database management system with custom script language handling 400+ scan curves from 41 detectors
• Proposed and implemented RANSAC-based (RANdom SAmple Consensus) curve fitting and linear interpolation to achieve noise-robustness and better resolution in curve analysis
• Designed and trained conditional autoencoder to predict detector response across temperature, humidity, and bias voltage ranges
• Model predictions showed $R^2 = 0.99$ on linear correlation between breakdown voltage and temperature
• Model reconstructs training curves with $RMSE = 0.090$ and showed a low $4.9\%$ slope error compared to training data
• Research presented on CERN Detector Research and Development Working Group 2 (CERN DRD3 WG2) Meeting in October 2025, and Coordinating Panel for Advanced Detectors (CPAD 2025) at Penn

Tech Stack: Python, PyTorch, RANSAC

DJMAX Rhythm Game AI Player

Term Project for Computer Vision | Team: Me, Jonie Nishimura, Tiffay Gao. | May 2025

Built a real-time AI player for DJMAX Respect V, a famous rhythm game, using deep learning in 2 weeks. The system outstrips average human performance, consistently scoring above 97% accuracy.

My Contributions:

• Collected and curated 139,500 training frames through gameplay
• Customized ResNet18 + GRU backbone with modified pooling layers and track-specific feature map splitting to predict multi-track keypresses in the game
• Designed a dynamic loss function with moving average smoothing to model timing tolerance, transient emphasis to prioritize keypress predictions, and class imbalance handling for robust training
• Optimized model to run at ~45 FPS in real time with 97% average accuracy

Demo Video:

This video shows the AI playing the song "L" in real-time on a MacBook. The AI triggers the keys automatically without human input, consistently catching fast paced notes with high accuracy, and breaks my personal record at the end.

More demos can be found on Google Drive

Tech Stack: Python, PyTorch, OpenCV, NumPy

Team Roles:

Lixing Wang (Me): Model architecture & loss function design, data collection pipeline (capture.py), dataset implementation (dataset.py), inference optimization, real-time gameplay testing.
Tiffany Gao: Model training & evaluation, core model implementation (model.py, train.py, test.py), video dataset class design, visualization & graphing.
Jonie Nishimura: Preprocessing functions (validation, trim_video), inference pipeline (play.py), TensorFlow-to-PyTorch conversion, poster design.