Choose timezone
Your profile timezone:
XCC - weekly meeting
→
America/Los_Angeles
Description
Meeting summary
Quick recap
This meeting focused on preparing for an upcoming drawing calibration workshop and reviewing progress on jet calibration studies. Michael presented initial draft slides for the workshop, showing performance differences between Z+jets and QCD samples using VBF selections, with particular emphasis on the degradation of performance in Z+jets due to fewer forward tracks. The team discussed technical issues with RPT calculations, noting discrepancies between Michael's approach and the established method used by Wasiko and Umar, which includes a delta R cone cut of less than 0.2. Ariel emphasized the need to understand why performance differs between ITK-only and Z+jets samples before proceeding with HGTD studies, suggesting that the current waves approach may create circular dependencies by using the same tracks for both vertex T0 determination and jet identification. Umar and Amir presented promising results from their jet energy regression work using neural networks, showing significant improvements in jet energy resolution and correlation with truth values, though Ariel advised caution in how resolution metrics are defined and presented. The team also discussed data storage optimization for their showering studies in ITK, with Michael reporting that the full TTBAR pileup truth sample required 500 gigabytes of storage and suggesting the use of clusters rather than individual hits as a potential compression strategy.
Next steps
Amir
- Work with Umar to incorporate cells into the jet energy regression model, possibly using a transformer architecture with separate encoders for cells and tracks.
Benjamin
- Resume work on showering in ITK after completing other commitments.
Michael
- Create a single Google Doc for the Alternative Calibration Workshop abstracts, add his own abstract, and share the link via email to the group mailing list, reminding everyone to upload their abstracts.
- Sync with Wasikul and Umar on the correct definition and calculation of RPT, including the Delta R < 0.2 cut.
- Remake the Z plus jets performance plots using the correct dilepton selection (electrons and muons only, no hadronic taus) and ensure the response is centered at 1.
- Modify the event displays to annotate which jets are used in the VBF selection.
- Investigate and potentially change the VBF jet selection logic (e.g., highest PT jets first, then apply Delta Eta cut).
- Study Z plus jets events based on the number of forward jets (one vs. two) to understand where HGTD can help.
- Prioritize events where HGTD can help in the analysis and determine the fraction of such events.
- Look into storing clusters (space points) instead of individual hits in the n-tuples to reduce file size and align with tracking concepts.
Umar
- Make differential plots (in PT and ETA bins) for the jet energy regression study.
- Investigate the jet energy resolution using the standard Gaussian fit method to extract stochastic and constant terms.
- Try training the jet energy regression model with and without tracks to disentangle their effects.
- Test the trained model on different jet flavors (e.g., b-jets) to study performance and potential biases.
- Start looking into JVT2 and provide updates.
Summary
Drawing Calibration Workshop Preparation
The team discussed preparations for an upcoming drawing calibration workshop. Michael uploaded a rough draft of materials based on previous global performance meeting slides, including additional event displays. Ariel suggested reviewing the content rather than the presentation format, and the group agreed that Michael would create a Google Doc to collect abstracts from all participants. Michael volunteered to email the mailing list to remind everyone to upload their abstracts to the shared document.
VBF Selection Performance Issues
The team discussed performance issues with the VBF selection in Z plus jets samples, focusing on jet separation criteria. Michael presented event displays and performance metrics showing a significant drop in purity fraction from 90% to 58% compared to the standard Higgs to Invisible analysis. The group debated whether to loosen the current delta-eta cut of 3.0, with Aram suggesting MJJ could be used instead, though Ariel noted that MJJ cuts are typically high (around 80 GeV) and PT-dependent. Michael mentioned statistics challenges with only 60,000 unweighted events out of 15 million total, and the discussion was cut off as he was about to show additional performance plots.
Track Jet Association Performance Issues
The team discussed performance issues with track jet association, particularly focusing on differences between C++ jets and QCD events. Michael presented plots showing degraded performance for forward track scattering tracks and RPT selection, which Ariel and Aram noted was expected due to the different axis ranges shown. The team identified a need to align on the proper method for calculating RPT, specifically incorporating a delta R cut less than 0.2 between ghost association and jet cone methods as previously discovered. They agreed to investigate the performance differences between ITK-only and Z+jet events in both central and forward regions before proceeding with HTTD vertex T0 studies.
Jet Implementation Performance Discussion
Ariel explained that the current implementation using waves gives more weight to tracks in jets, which helps in BBF but creates issues in Zebra when only one jet is available. He suggested using an opposite approach by excluding tracks in the jet and considering orthogonal variables from event n-tuples. Michael agreed to follow up with Ariel and Umar to create a roadmap for fully characterizing the performance, including different jet scenarios. The team also discussed updating plots to properly handle leptonically decaying C events and avoid electronic tails, with Michael acknowledging that some plots were outdated from July.
HGTD Timing for VBF Jets
The team discussed challenges with using HGTD timing information for VBF jet identification, particularly in events with pileup jets that mimic hard scatter jets. Ariel suggested modifying the approach by excluding tracks within identified jets when calculating vertex T0, while still using waveforms for other tracks to determine jet timing. The group agreed to systematically analyze specific event displays to better understand when HGTD can help distinguish between pileup and hard scatter jets, with the goal of determining what fraction of events might benefit from timing information. Next steps include reviewing RPT with Wasco's code, examining Z-plus jets based on forward jet counts, and modifying event displays to better annotate which jets are used as VBF jets.
Jet Calibration Update Discussion
Amir and Umar provided an update on their work on jet calibration, focusing on using tracks and cells to regress jet energy to truth. They discussed a two-sigma significance threshold and energy cut on cells, with Ariel confirming that this should not significantly impact their approach. Umar presented a feasibility study using a transformer architecture to regress jet response, operating on tracks associated with jets.
Neural Network Jet PT Correlation
Omar presented a 2D histogram comparing truth jet PT with reconstructed jet PT, showing that the neural network's predictions for jet PT have a stronger linear correlation with the truth values compared to the default reconstructed values. Ariel suggested exploring different eta regions to verify if the improvement holds across all regions, and noted that the network already incorporates jet eta information through a global token in the transformer. The discussion revealed some confusion about jet PT cuts, with Omar initially excluding jets above 30 GeV but later clarifying that the lower cut is around 17 GeV, though not all jets below 30 GeV are calibrated.
Response Measurement Methodology Correction
Ariel corrected Umar's understanding of response measurement, explaining that it should be centered at 1 rather than using median values, and emphasized the importance of using standard Gaussian methods for calculating resolution that can be published and understood by others. Ariel noted that while Umar's current approach provides interesting insights, it would be better to use established Gaussian fitting methods for measuring resolution components like stochastic and constant terms. The team agreed to iterate offline on the proper methodology, with Ariel offering to provide guidance on implementing the correct resolution calculations.
PT and ETA Analysis Discussion
Ariel and Umar discussed the PT and ETA ranges for their analysis, with Umar clarifying that EDA is inclusive while PT is greater than 30 GeV. Ariel suggested creating a matrix of plots showing distributions across different PT and ETA ranges, including differential plots and resolution as a function of PT in different ETA bins. They examined a 2D correlation plot showing the impact of the neural network, which significantly reduced the width and improved correlation. Ariel suggested comparing networks that use tracks versus calorimeter-only approaches to better understand the effects. They also discussed the importance of properly defining and presenting resolution metrics, with Ariel noting that the constant term should be 3% rather than 10%.
Parameterization and Jet Modeling Implementation
The team discussed implementing a parameterization with energy resolution including stochastic and constant terms, with Umar planning to use Gregor's previous work as a starting point. They agreed to explore training the model on different jet types (light vs. b-jets) and testing performance across various flavors to determine if inclusive training could work for all jet types. Michael reported running a full pileup truth sample with 50,000 events that resulted in 500 gigabytes of data, and the team discussed optimizing storage by using clusters instead of individual hits while maintaining the ability to show event displays.
There are minutes attached to this event.
Show them.
