Choose timezone
Your profile timezone:
CIDeR-ML General Meeting
→
America/Los_Angeles
Description
https://u-tokyo-ac-jp.zoom.us/j/83932834349
Recording
Minutes:
Quick recap
The meeting focused on Omar's presentation of differentiable simulation for large TPC detectors, where he demonstrated improvements to muon propagation modeling and energy loss calculations using CSDA (Continuous Slowing Down Approximation) for faster parallel processing. The team discussed technical details including Bragg peak relaxation parameters and gradient calculations for detector optimization. Patrick and the team also addressed ongoing challenges with 1-2% discrepancies in their Monte Carlo comparisons of muon energy loss and Cherenkov angle measurements in water. The conversation ended with status updates from team members including Yifan's work on segment fitting, Zheng's code merging and testing, and Riya's questions about the refactor implementation and PMT measurements.
Next steps
- Yifan: Continue fitting segment DDX position and calibration parameters, and follow up with Omar on Slack for further discussion about segment fitting.
- Riya: Clean up the wavelength branch and provide the relevant root file/data to Omar for integration and testing with the refactor.
- Omar: Continue working on the refactor, including merging Riya's SK branch and wavelength branch, and complete clean-ups to support both reconstruction and calibration.
- US team members: Review and update the Google Doc with individual "how-to-dos" for the upcoming workshop and report readiness in next week's meeting.
- Patrick: Determine new students' interests and inform Omar to help suggest suitable projects once interests are identified.
Summary
Differentiable Simulation for TPC Detectors
Omar presented on differentiable simulation for large TPC detectors, focusing on muon propagation and energy loss modeling. He demonstrated improvements to the implementation using the CSDA (Continuous Slowing Down Approximation) method, which significantly speeded up calculations from one second to 2 milliseconds. Omar also introduced a new loss function using a filter to improve gradient quality and presented a method to combine losses from different wire planes using a geometric mean approach with adaptive epsilon handling. The presentation included various gradient variation plots showing the effectiveness of the differentiable approach for muon parameters and detector parameters like diffusion and recombination.
Muon Tracking and Energy Estimation
The meeting focused on discussing a muon tracking and energy estimation method. Speaker 1 (SLAC-BLDG-41---Room-1154) explained how track length is determined based on muon energy, with energy-dependent range changes affecting the line patterns in their visualization. Yifan asked about segment fitting approaches, and Speaker 1 clarified that while they use segment-based optimization, the current presentation showed a line-based parameterization. The discussion also covered Bragg peak relaxation, where Speaker 1 indicated they used either one or two millimeter steps, noting that perfect peak fitting isn't critical for muons due to finite detector resolution. Finally, Speaker 1 explained the difference between serialization and parallelization approaches for energy calculation, describing how the parallelized method uses pre-calculated energy deposition integrals to sample energy loss based on distance.
Muon Energy Loss Modeling Discussion
The team discussed discrepancies in their muon energy loss modeling in water, particularly regarding the collapse of Cherenkov rings. Patrick explained they are seeing about 1-2% disagreement between their data and Monte Carlo simulations, which they suspect may be related to energy loss modeling rather than the Cherenkov process itself. The group also reviewed individual project updates, including Zhenxiong's work on merging code and testing sampling methods, Yifan's fitting of DDX position segments, and Riya's wavelength analysis work. The team discussed the upcoming workshop preparation, with a focus on US participants needing to report their readiness next week and review the shared Google document containing individual to-dos.
