Forward Tracking in the ILD Detector

ILC

the goal Standalone track search for the FTDs (Forward Tracking Disks)

Digitizer Vertextion Reconstruction Track Reconstruction

Why not combine with TPC tracking?

TPC

FTDs

Methods Cellular Automaton Kalman Filter Hopfield Neural Network

The Cellular Automaton

It's about rules

IP

cell ( segment)

state 0

On the FTDs

The Hits

The Segments

Cellular Automaton

Clean Up

Longer Segments

Cellular Automaton

Clean Up

Track Candidates

Kalman Filter Prediction → Filtering (Updating) → Smoothing Superior to simple helix fitter Quality Indicator: χ ² probability

Kalman Filter Prediction → Filtering (Updating) → Smoothing Superior to simple helix fitter Quality Indicator: χ ² probability

Track Candidates

Kalman Filter

p > 0.005

The Hopfield Neural Network Track ↔ Neuron Goal: the global minimum

T=2 T=1 T=0

Background Inner 2 disks are pixel detectors How many bunch crossings? 100 BX * hit density ≈ 900 hits / pixel disk

Ghost hits Outer 5 disks are strip detectors Solution: shallow angle

At the moment Conversion to new geometry (staggered petals) Debugging and efficiency Seperating core form implementation Sensible use of Hopfield Neural Network Robustification Picking up hits from other places More analysis Individual steering for pattern recognition → xml file Measure the improvement (old vs. new software)

Thanks to Rudi Frühwirth and Winfried Mitaroff Steve Aplin, Frank Gaede and Jan Engels And Jakob Lettenbichler (Belle 2)

Thank you!

Back Up

Dependencies FTD drivers and gear → at the moment in between solution The real background Bad χ ² probability distribution Number of integrated bunchcrossings

Before Hopfiel Neural Network

After Hopfield Neural Network

tracks will Skip a layer Connect directly to the IP

regions