Focus on: 
Login

CIDeR-ML General Meeting

→ America/Los_Angeles
Description

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

Link to Recording

Minutes:

Quick recap

The meeting focused on updates from two main presenters: Zhenxiong and Nahuel. Zhenxiong reported on SIREN simulation improvements, including updating PMT saturation from 30 to 70 to better handle high-charge events and separating training data from reconstruction data to improve accuracy. The team discussed the need to check for dead PMTs and investigate unusual 2D plots in the data. Patrick mentioned an upcoming WCTE collaboration meeting on September 7 and plans to invite Zhenxiong to present this work at future analysis meetings. Nahuel presented progress on learning water absorption variations in SuperK detector using Lucid, showing how simulated data can match real PMT data by adjusting reflection rates and other parameters. The team discussed challenges with fitting the absorption field and noted discrepancies between the simulated and actual data, particularly in the top cap and first row PMT histograms.

Next steps

Nahuel

  • Investigate the negative values in the recovered absorption field and adjust the field definition in Lucid to ensure positivity.
  • Further investigate the misalignment of peaks in the top histograms and the effect of the refractive index parameter on the distribution.
  • Riya: Continue running SK simulation events now that it is working for SK4.

Patrick,

  • Invite Zhenxiong to present his work at an upcoming WCTE analysis meeting before the collaboration meeting on September 7.

Zhenxiong

  • Double check the PMT masking in SIREN and ensure that non-working PMTs are properly ignored.
  • Create event displays for SIREN predictions to visualize PMTs that are off.
  • Show how the predicted charge per PMT changes from pre-tune to post-tune in SIREN, and compare with known bad PMTs in real data (Ryotaro).
  • Investigate the cause of the large "blob" in the 2D plot and the lack of improvement in PE resolution post-tuning.
  • Check the event displays (average over all events) to investigate the zero bin issue, especially in Monte Carlo.
  • Process the raw data from Ryotaro to be readable by SIREN and perform reconstruction with pre-tune and post-tune.
  • Check the repository or with Ka Ming for the input files used in fiTQun to potentially rerun and output per-PMT PE.

Summary

PMT Saturation and Reconstruction Updates

Zhenxiong updated the PMT saturation setting from 30 to 70 to ensure all charge contributions are included in the loss computation. She also separated the reconstruction data from the training data to improve accuracy and found that post-tune results are now closer to the fiTQun results. Zhenxiong plans to process raw data from Ryotaro and perform reconstruction with pre-tune and post-tune settings. Patrick suggested checking PMT efficiency variations and creating an event display to identify non-functional PMTs for cross-validation with known detector issues.

PMT Performance and SIREN Issues

The team discussed issues with PMT performance and SIREN simulation results, particularly regarding unusual bright bins in both real data and MC simulations. SLAC suggested checking if the problem stems from SIREN mispredicting charges and recommended applying energy cuts to improve the data fit. Patrick mentioned that while reconstructed quantities were shown in the next slide, the per-PMT information from fiTQun might need to be rerun, as Zhenxiong only had the .h5 files without the full input details. The group also noted that the vertex resolution didn't improve significantly with post-tuning, and Junjie suggested investigating whether this is due to voxel size issues or other factors.

Water Absorption Detection Variations Study

Nahuel discussed his work on learning variations of water absorption in detectors like SuperK or HyperK using Lucid. He explained the process of comparing simulated PMT data to real PMT data, using a siren grid instead of a full siren to address memory errors and training time issues. The siren grid is interpolated in a 12x12x12 grid to improve performance.

SIREN Model Calibration Progress

Nahuel presented progress on using Super-K calibration data to train the SIREN model, demonstrating how reflection rates affect histogram patterns and showing initial fits of data points from the SK Calibration paper. He noted that while the model captures the general trends correctly, there are still issues with peak alignment in the top histograms and negative absorption field corrections that need addressing. Nahuel plans to investigate these discrepancies further, particularly regarding the laser position parameters and the definition of the absorption field in Lucid.

Lucid Model Discrepancy Discussion

The team discussed discrepancies between the Lucid model and the SK model, with Riya explaining that physics implementation differences, including potential reflection and QE issues, likely cause the observed differences. Patrick noted that the SK constants from the paper don't match out of the box, and Riya confirmed she is working on running SK simulation MC for SK4 to generate more events. The team identified a need to examine how the refractive index affects the distribution, similar to how they analyzed the reflection parameter.