Login

CIDeR-ML General Meeting

→ America/Los_Angeles
Description

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

Link to recording

 

Quick recap

This meeting focused on updates and discussions regarding simulation improvements for the SK detector using LUCiD software. Riya presented her work on creating a faster simulation method that reproduces SKDETSIM results, showing good agreement between LUCiD and SKDETSIM in terms of photoelectron predictions and event displays. She identified a PMT geometry mismatch between LUCiD's spherical model and SKDETSIM's more complex donut-shaped model, which affects time-of-flight comparisons, and discussed plans to implement proper PMT geometry modeling. Omar presented multiple fixes for simulation issues, including improvements to the SIREN energy parameterization using a degree 7 polynomial fit in log space, corrections to photon deposition on sensors behind walls, and adjustments to overlap calculations. He also discussed ongoing work on implementing full 3D geometry in LUCiD, which would enable extensions to other detector configurations like Juno. Junjie reported on hyperparameter optimization for their model, finding that model depth impacts visibility and T0 prediction performance. The team discussed plans for the upcoming Doraemon meeting in Kyushu, where they will listen to a presentation at 9:00.

Next steps

Riya

  • Investigate and fix the PMT geometry mismodeling issue in LUCiD, considering the effect of reflections and gradients. Consult with Omar for the best approach.
  • Repeat the tests with Mie scattering and reflection physics enabled, while disabling other water physics.
  • Begin to include gradient checks in the simulation tests.

Junjie

  • Finalize the training of the model for the paper and document the findings and limitations.

Omar

  • Push the code with the fixes (fixes 1 and 2) to the main branch of LUCiD and notify the team in Slack when it's done.
  • Continue working on the full 3D geometry implementation for LUCiD (LUCiD 2.0) and keep the team updated.

Summary

LUCiD Simulation Method Updates

Riya presented updates on her work to create a faster simulation method with LUCiD that reproduces SKDETSIM results, including matching laser emission models and implementing proper reflection models. She demonstrated that the number of predicted photoelectrons and event displays between LUCiD and SKDETSIM were very similar, with small charge differences per PMT. Riya also discussed time-of-flight comparisons, noting some differences due to a Gaussian timing mixture in SKDETSIM based on PMT photon hit timing structure.

PMT Geometry Model Discrepancy

Riya presented her work on modeling photon detection in a detector simulation, identifying a significant difference between LUCiD's spherical PMT model and SKDETSIM's more complex donut-shaped model, particularly in the top cap region. She implemented a temporary fix using weighted acceptance based on arrival angles but acknowledged this approach may not be optimal. The team discussed that properly addressing the PMT geometry mismatch would require either changing the shape model or implementing a custom PMT model, with considerations for how different geometric models might affect gradient calculations and backward path functionality in the simulation.

PMT Simulation Gradient Checks

The team discussed implementing gradient checks and geometric modeling for PMT simulations. Omar suggested using an imaginary spherical enclosure around PMTs to perform proper checks, which would make simulations slightly slower but improve gradient accuracy. Riya agreed to focus next on implementing reflections, acknowledging that changing the PMT geometry would require modifying the reflection model. The group noted that while the current empirical fit approach may handle small effects, it's important to properly assess the impact of these geometric changes before determining their significance.

SIREN Energy Parameterization Updates

Omar presented changes to the SIREN energy parameterization, explaining that the original power law approach had mismatches at low energies that created energy bias in reconstruction. They replaced the power law with a degree 7 polynomial in log space, achieving errors on the order of 0.1% and improving reconstruction results for the first few energy bins. The next steps include putting the updated parameterization on GitHub and eventually integrating it into the SIREN file process.

Simulation Physics Challenges and Fixes

Omar discussed challenges with multiple contradictory effects in their simulation, particularly focusing on a physics difference where photons were redepositing into PMTs after reflection, causing about 5-6% charge change. This issue was identified and fixed. The team also addressed a second case involving photon hits at grazing incidents that could contribute to multiple sensors, which was capped to one contribution. Additionally, Omar mentioned a small charge bias from sim scaling that required adjusting the overlap width and changing the calculation to use float64 for the overlap table.

LUCiD Photon Deposition Fixes

Omar presented fixes for photon deposition issues in LUCiD, categorizing them into two groups with Fix 2 being more problematic due to gradient issues. The fixes were shown to improve momentum unbiasedness but slightly worsen position and T0 predictions. Omar indicated the code would be pushed to the main branch that day and confirmed it would be safe for Riya to use for upcoming tests. Omar also mentioned ongoing work on full 3D geometry implementation, which could take a few weeks to complete, and discussed plans to help implement LUCiD for JUNO with actual 3D geometry.