CIDeR-ML General Meeting

Thursday 3 Sept 2026, 17:00 → 18:00 America/Los_Angeles

https://u-tokyo-ac-jp.zoom.us/j/83932834349

Link to Recording

Quick recap

This meeting focused on project updates and validation of sirenTV, with Junjie presenting the main findings. Junjie reported significant improvements using a Poisson log likelihood ratio loss function, which successfully addressed fluctuation issues in low visibility regions by giving more importance to high visibility regions. The new loss function achieved better performance with half the training time compared to the previous Huber loss approach. Junjie also discussed challenges with PCA component supervision and explored reconstructing quantiles from PCA, though this approach didn't yield optimal results. Omar provided updates on LUCiD development, including bug fixes for secondary light contributions and ongoing work on hyperparameter optimization for the Gauss Newton optimizer. He also mentioned progress on a new 3D geometry implementation that can model detailed PMT structures including MPMTs and covers. Ryo shared that he received PE function files from John and is working on implementing them, with amplitude normalization still pending. The conversation ended with a discussion about potentially combining multiple SIRENs for different detector components, with Kazuhiro suggesting they should review previous documentation and considerations before proceeding with new discussions.

Next steps

Junjie

• Investigate further the weighting scheme in the probability density domain for PCA components and continue ablation tests.
• Explore a combination of loss functions for visibility to improve performance in both high and low visibility regions.
• Continue working on the training with supervision on gradients for sirenTV with waveform.

Omar A. Alterkait

• Finish rerunning the calibration and reconstruction scripts from the paper to ensure everything is working, then push the code fix for the secondary light bug.
• Consider automating tests for the LUCiD code in CI, taking into account GPU requirements.

Patrick

• Start a Google document to lay out the ideas and challenges for linking multiple SIRENs, based on previous discussions.

Ryo

• Complete the check of the PE function received from John, then implement it in the simulation software.

Zhenxiong

• Finish preparing the slides and address comments by the end of the week, and present remotely next week.

Summary

sirenTV Loss Function Ablation

Junjie presented the results of an ablation test comparing different loss functions for sirenTV, demonstrating that the new Poisson log likelihood ratio loss significantly improved performance in high-visibility regions compared to the previous smooth Huber loss. The new loss achieved better results with half the training time and reduced spuRyous fluctuations in low-visibility regions. Junjie also discussed challenges with PCA component supervision and explored reconstructing quantiles from PCA, though this approach showed mixed results and requires further investigation.

opticSIREN Bias Discussion

Patrick asked Junjie about a previous bias issue in the low visibility region for the water Cherenkov opticSIREN that was seen three years ago. Junjie explained that he had used L2 loss for opticSIREN but didn't recall trying negative likelihood ratio loss. Junjie noted that opticSIREN has limitations on high visibility and suggested that the new approach, which emphasizes the high end and uses a ratio in loss, should help address both the high and low visibility issues.

LUCiD Development and Bug Fixes

Omar discussed ongoing work on LUCiD, including code fixes for a bug with secondary light contributing to sensors and hyperparameter tuning for the Gauss Newton optimizer used in reconstruction calibration. Omar explained that while some testing automation exists in the main branch, fully running the paper analysis requires significant GPU resources and would need to be carefully implemented in CI. Patrick mentioned that the Super-K remains stalled until Riya demonstrates more progress and need.

SKDETSIM and WCTE Progress Updates

Patrick discussed progress on comparing SKDETSIM analysis with LUCiD, noting limited advancement due to summer travel and other commitments for Riya. He mentioned that the MOU for WCTE was sent out for review with a deadline today for comments. Omar shared updates on a new 3D geometry mode he's developing for LUCiD, which can model detailed PMT structures including MPMTs and covers, though gradient calculations still need refinement for calibration and reconstruction purposes.

PE Function Implementation Progress Update

Ryo reported receiving a PE function from John to implement simulation, which he is currently reviewing. He noted that the software implementation is nearly complete except for amplitude normalization, which will be determined through calibration. Patrick suggested considering implementation in LUCiD for future development.

opticSIREN Calibration Presentation Update

Zhenxiong confirmed she will prepare slides addressing comments and present remotely next week, with Lawrence offering to schedule her later. Patrick reported WCTE's comments on opticSIREN cosmic calibration work, noting that while initially difficult to grasp, people became interested in the potential applications and modeling capabilities of SIRENs. The discussion concluded with a reference to a previous idea about linking multiple SIRENs for different detector components, though this was previously tabled due to complexity.

Silent Features Combination Discussion

Patrick and Kazuhiro discussed the feasibility of combining additional SIRENs, with Kazuhiro suggesting they should first review previous discussions rather than starting from scratch. Kazuhiro recommended creating a Google document to outline ideas and trade-offs, particularly regarding data constraints for model optimization. Patrick Tsang noted that tools like LUCiD were not available during previous discussions two years ago, suggesting they should reconsider the approach based on current capabilities.