CIDeR-ML General Meeting

→ America/Los_Angeles
Description

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

Recording

Minutes:

Quick recap

The meeting focused on technical progress updates and discussions across multiple projects. Riya presented on implementing physics models from SKG4 into LUCiD, including various scattering mechanisms like Rayleigh and Mie scattering, though she noted challenges with differentiability and combining scattering lengths. Zhenxiong shared validation work on OpticSIREN performance, comparing predicted charges with Geant4 simulations and showing improvements after tuning with true track information rather than reconstructed information. The team discussed concerns about reconstruction biases and the need for better modeling of the OpticSIREN. Other participants provided brief updates, with Sam reporting fixes to DDP training issues for SirenTV, and Yifan noting ongoing work on the need for reconstruction improvements. The conversation ended with discussions about hotel arrangements for an upcoming Tokyo trip, with plans to determine prioritized room bookings.

Next steps

  • Riya: Implement Raman scattering in LUCiD in the future (not immediate, but noted as a longer-term task).
  • Riya: Look into inverse CDF methods and the reparameterization trick to enable differentiability for g in the Henier-Greenstein phase function, and follow up with Omar offline for further clarification.
  • Zhenxiong: Double check the lower bin for the ratio plot and update plots as needed to ensure improvement in bias/spread is correctly shown.
  • Zhenxiong: After current training finishes, update plots with the latest results, including loss curves, information on training duration, and number of events used.
  • Zhenxiong: Use different line styles or marker styles in plots to facilitate comparison, so same color can be used to eash comparisons.
  • Zhenxiong: Check 68% intervals for reconstruction-true distributions to directly compare to chosen smearing values.
  • Zhenxiong: Overlay the fiTQun curve with the 100 and 200 smearing curves to directly compare performance.
  • Zhenxiong: Perform analysis using the OpticSIREN visibility map with point source (shotgun sample) instead of cosmic sample to check for bias in the OpticSIREN model.
  • Zhenxiong: Investigate the effect of changing the sampling scheme (e.g., binning size) on results and report if significant changes are observed.
  • All (or relevant team members): Individually confirm to Patrick via Slack if they need hotel bookings for the April Tokyo trip, and provide prioritized list if needed.
  • All: Notify the travel office of any hotel changes if after DOE approval.

Summary

Physics Models Implementation in LUCiD

Riya presented her progress on implementing physics models into LUCiD, focusing on scattering processes. She explained the differences between the current LUCiD implementation and SKG4, particularly in the handling of Rayleigh and Mie scattering. Riya also discussed the challenges of implementing Raman scattering and the need to determine the overall scattering length in LUCiD. She sought Omar's help in addressing these issues, particularly regarding the non-differentiable nature of the current implementation.
Omar and Riya discussed the implementation of scattering, focusing on the sampling method used for interaction lengths. Omar suggested using probabilities for sampling instead of the SKG4 minimum length approach. They also talked about the potential future inclusion of Raman scattering, but agreed it should be left for later, since SK doesn't have that low wavelength laser calibration. Omar advised Riya to look into inverse CDF methods and reparameterization for improving gradient calculations in the simulation. They decided to follow up offline on this topic as Riya had more questions.

OpticSIREN Performance Validation

The team discussed validation of OpticSIREN performance, comparing predicted charge with Geant4 simulated true PE. Zhenxiong presented results showing improved tuning when using true track information instead of reconstructed information, with reduced bias and uncertainty in the opticSIREN predicted charge. Patrick de noted that the improvement in the zeroth bin was not clearly shown in the plots, and Zhenxiong agreed to double-check this. The team also discussed the impact of smearing on position and direction, with Zhenxiong presenting plots showing how uncertainty increases with larger smearing values.

Enhancing Training and Performance Metrics

Patrick de and Zhenxiong discussed the need for more detailed information in the training process, including loss curves, duration, and number of events. They agreed on improving plot styles for better comparison and calculating 68% quantiles to know where fiTQun lies between the chosen smearing values. Zhenxiong acknowledged the similarity between the fiTQun curve and the 100 and 200 curves, and Patrick de suggested overlaying them for clarity. They also discussed the need to ensure performance metrics meet requirements before moving on to data usage, with a focus on improving reconstruction in the top-right quadrant. Patrick proposed using the OpticSIREN for simultaneous reconstruction and tuning, which Zhenxiong confirmed could be feasible.

OpticSIREN Simulation Analysis

The team discussed performance comparisons between OpticSIREN and Monte Carlo simulations, with Patrick Tsang raising questions about bias and model accuracy. Zhenxiong agreed to prioritize further analysis using pre-tuned samples and investigate potential sampling scheme adjustments.
The group also addressed issues with distributed training for SirenTV, which Sam reported was now fixed and in progress. Additionally, they discussed hotel arrangements for an upcoming Tokyo trip, with Patrick requesting participants to provide their names for booking purposes, noting limited room availability.