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
Yiyang presented on electron tracking with simulation-based inference and probabilistic programming language in a toy chain contactor model. Yiyang discussed the challenges of reconstructing low-energy electrons in water due to severe multiple scattering, explaining how traditional methods using marginalized PDFs may lose important covariance information among photon tracks. He presented two approaches: simulation-based inference using deep sets to approximate likelihood ratio corrections, and probabilistic programming language implementation that translates Gen4 physics processes into a probabilistic framework. Both methods found that while photon direction information contains measurable non-IID structure beyond marginalized profiles, it does not significantly improve initial electron angular resolution due to degeneracy in track reconstructions. The work did demonstrate that scattering severity can be measured and used for event classification, and timing information provides stronger constraints on electron directions when PMT-TDS precision is sufficient. Junjie suggested Yiyang present more detailed breakdowns of the methods in future sessions to better digest the complex technical content.
Next steps
Yiyang
- Prepare and present a more detailed presentation on one of the methods (simulation-based inference or probabilistic programming language) at the next meetings.
Summary
Electron Tracking Simulation Research
Yiyang presented a study on electron tracking with simulation-based inference and probabilistic programming language in a toy chain contactor model. The research aimed to improve angular resolution for low-energy electrons in water by leveraging information beyond marginalized electron tracks, as traditional reconstructions lose covariance among tracks. Yiyang demonstrated that while simulation-based inference could recover track-aware likelihood and classify events based on multiple scattering severity, it did not significantly improve angular resolution itself, though the log-likelihood ratio could be used to indirectly boost resolution by identifying more track-like, direct events.
P0 vs P-TRACK Comparison Discussion
Yiyang explained the difference between P0 and P-TRACK, noting that P0 uses a manual template to multiply marginal PDFs while P-TRACK requires numerical approximation through a classifier trained on both ID and Gen4 track samples. Junjie questioned how the photon propagation data is generalized into a likelihood distribution, and Yiyang clarified that a shared photon encoder processes variable photon numbers into fixed-dimensional features before classification. The discussion revealed that their toy model currently only tracks electron scattering without light transportation, focusing on electron scattering at low energies.
Probabilistic Programming for Geant4 Physics
Yiyang discussed the challenges of implementing probabilistic programming for Geant4 physics processes, including variable length dimension inference and the use of fixed track segment numbers with varying track lengths to improve efficiency. They calibrated the model using Geant4 input as a benchmark and found that while photon directions provided some measurable non-IID structure, they did not significantly improve initial electron angular resolution due to degeneracy. Yiyang concluded that timing information could provide stronger constraints on electron initial direction but requires precise PMT-TTS, and suggested future improvements including cross-energy calibration and using classifiers instead of marginal templates to train the PPL model.