Choose timezone
Your profile timezone:
CIDeR-ML General Meeting
→
America/Los_Angeles
Description
https://u-tokyo-ac-jp.zoom.us/j/83932834349
Link to recording
Quick recap
The meeting focused on updates from the team on different projects related to waveform modeling and simulation. Junjie presented detailed results on sirenTV training and evaluation, discussing its performance compared to PCA and lookup tables, and highlighting its strengths and limitations in predicting PMT waveforms. He also shared findings on the differentiability and computational efficiency of sirenTV, as well as plans for further training. Patrick, Naomi, and Riya reported on their participation in the HyperK collaboration meeting, with Naomi noting progress in implementing light injection sources with Lucid. Yiyang and Zhenxiong provided brief updates on their ongoing work, with Zhenxiong seeking guidance on paper preparation. Sam clarified the roles of Chroma and sirenTV in the simulation pipeline. The conversation ended with a round of updates from all participants.
Next steps
Junjie
- Perform another training for Siren using full resolution quantile bins.
- Align PMT waveforms in one event during training to see if it improves learning.
- Complete the paper by the end of October after the new training is finished.
Naomi
- Give a presentation next week on the Hyper-K collaboration meeting and Lucid-based approaches.
Riya
- Give a presentation next week on the Hyper-K collaboration meeting.
Yiyang
- Show more details about the probabilistic programming differentiable electron model in the next week or the week after.
Zhenxiong
- Continue running reconstruction for beam samples with post-tune and pre-tune Siren and overlay the results.
- Decide whether to start writing the paper based on the 2-5% risk regarding the MOU approval.
Summary
Siren Training Performance Results
Junjie presented results from Siren training and evaluation, showing good performance in predicting PMT waveforms with biases below 1% for visibility and T0, though quantile predictions had larger biases. The model achieved sub-voxel position resolution of around 6 millimeters when using stochastic sampling methods, though it performed slightly worse than PCA in optimization tests. Junjie identified that the main source of bias in Siren came from the reduced resolution quantile input rather than visibility predictions, and planned to conduct additional training with full resolution quantile bins as a next step.
Waveform Modeling Implementation Discussion
Patrick discussed technical challenges with understanding the waveform modeling approach, and Junjie explained that a waveform lookup table would be needed for Cherenkov applications. Junjie outlined two implementation options: using a fast simulator similar to Sam's work on Chroma, or using Siren as a surrogate model. Sam clarified the differences between Chroma, Siren, and Goop, explaining that Chroma provides full detector simulation while Siren focuses on PMT response, and emphasized that Siren's differentiable nature makes it preferable for calibration purposes. Junjie indicated the plan is to complete one more week of training and then finalize the work by the end of October.
Project Updates and Progress Reports
The team provided updates on their recent work and ongoing projects. Naomi reported progress on implementing light injection sources using Lucid with Hyper-K geometry and mentioned having a presentation ready for next week. Riya confirmed she was preparing slides for the HyperK collaboration meeting and would present next week as well. Yiyang discussed his transition from Tsinghua to ICRR and planned to share details about probabilistic programming and a differentiable electron model in the coming weeks. Zhenxiong reported running reconstruction work on beam samples and sought guidance about whether to proceed with paper writing while waiting for MOU approval, with Patrick suggesting it was likely safe to continue with only a small risk of issues.