Login

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 and discussions related to the development and application of SIREN models and other reconstruction algorithms for WCTE data. Omar A. Alterkait reported on ongoing bug fixes and rerunning of the reconstruction and calibration pipeline in LUCiD, noting some issues with photon contributions and edge cases. Zhenxiong presented preliminary results on beam reconstruction using WCT data, comparing the performance of SIREN with other methods and outlining a draft structure for a paper on SIREN applications in WCTE. Junjie shared detailed findings on the limitations of SIREN models in learning high-frequency features, particularly near PMT surfaces, and discussed ablation studies on model architecture and omega zero settings. Yiyang briefly introduced a project on electron tracking using simulation-based inference, highlighting potential improvements in angular resolution.

Next steps

Junjie

  • Finish the depth ablation for the SIREN model and determine the best combination of omega_0 and depth for each target.

Omar A. Alterkait

  • Fix bugs in LUCiD, rerun the reconstruction and calibration pipeline, and present results hopefully next week.

Patrick

  • Follow up on the MOU review status regarding the use of cosmic muon data for the SIREN paper.

Yiyang

  • Send slides to Patrick or Kazuhiro and schedule a dedicated presentation for two weeks later.

Zhenxiong

  • Try the pre-tuned SIREN model for beam reconstruction and update the analysis accordingly.
  • Finalize the draft of the SIREN paper for WCTE, aiming for completion in one month, and upload the outline link to the Indico page.

Summary

Beam Reconstruction Progress Update

Omar A. Alterkait reported delays in presenting his work due to GPU issues but explained he is fixing bugs in LUCiD and rerunning the reconstruction and calibration pipeline. The main bug identified affects photon contribution to sensors, changing light amounts by 4%. Zhenxiong presented preliminary beam reconstruction results for WCTE using data from run 1478, comparing different reconstruction methods including OpticSIREN and fiTQun. The results showed OpticSIREN achieved higher resolution than fiTQun, particularly in momentum reconstruction, though some data selection improvements may be needed.

Bimodal Structure and Data Reconstruction

Patrick and Zhenxiong discussed the bimodal two-peak structure in vertical direction plots, with Zhenxiong noting uncertainty about whether this was due to statistics and planning to explore it further. Ryo provided information about the current status of using fiTQun and LicketyFit algorithms for data reconstruction, indicating some people are testing these but the implementation details are not fully established. Zhenxiong presented a draft outline for a SIREN paper for WCTE, which will cover the demonstration of the SIREN model, WCTE, calibration and reconstruction using WCTE data, including sections on the SIREN model, WCTE detector, cosmic muon and beam selection, and validation results.

Paper Draft and MOU Review

The team discussed a paper draft that Zhenxiong estimated would take about one month to complete. Junjie inquired about whether the paper would be a WCTE collaboration paper, but Patrick clarified it would be a technical paper with additional authorship. Patrick confirmed that the MOU review deadline had passed two weeks ago with no comments, indicating approval to use the cosmic muon data.

SIREN Prediction Performance Analysis

Junjie presented findings on SIREN prediction performance limitations for a toy detector, analyzing frequency domain characteristics of different targets including visibility, T0, and PMT waveforms. The analysis revealed that PMT surfaces contribute significantly to high-frequency features, with visibility showing the strongest high-frequency content followed by PCA coefficients and T0. Junjie conducted ablation studies comparing different model architectures and training approaches, finding that separating T0 and visibility targets improved T0 prediction by 2-3 times, and that larger models with appropriate omega0 values are needed for PCA coefficient training due to higher frequency content.The team discussed how to select appropriate omega0 values for real data calibration, with Junjie explaining that omega0 is primarily determined by the type of data being processed rather than specific calibration needs. Patrick and Junjie agreed that model depth has a more significant impact on handling high-frequency features than omega0, which controls convergence speed.
Yiyang briefly presented on electron tracking with simulation-based inference, discussing approaches to improve angular resolution beyond marginalized general photo angle PDF methods, though the full presentation was deferred to a future meeting due to time constraints.