Login

EP R&D Software Working Group Meeting

→ Europe/Zurich
Vidyo

Vidyo
Live Notes
EP R&D Software Working Group Meeting
Zoom Meeting ID
62893115044
Host
Graeme A Stewart
Alternative host
Andre Sailer
Useful links
Join via phone
Zoom URL
Hide

Software R&D Working Meeting Minutes
====

Introduction
---

Next meetings - Erica will check for a good date for calorimetry with Felice and Marco.

Plans for 2021
---

### Calo

CLUE is nearest neighbour algorithm (ref. <https://www.frontiersin.org/articles/10.3389/fdata.2020.591315/full>).

A lot of efforts on physics performance, but also performance is taken into account and monitored.

### Tracking

Finalising first ACTS paper.

ML finding was quiet similar to when people tried to use this for ATLAS EMEC geomentry.

vecmem - anticipate CAF presentation, this is a very common problem and not generally considred to be solved yet.

Faster Simulation
---

Inference numbers are for single particles. Geometry is a PbWO4 barrel calorimeter. If one trained on a more complex geometry the full simulation time would be even higher.

LWTNN has to deal with JSON numbers, so less type information. This drives up the runtime memory consumption.

Loss of information at deep r values (slide 22) could be due to lack of training data. One would then retrain addressing this defect. N.B. one network, one training, with energy conditioning, which explains why the defect is seen at all energy values.

Have you looked at incident angles? A. No, not yet. Is it worth optimising then, when this is such an important point? ONNX optimisations are quite generic, so should apply in all cases (including ones with incident angle). Incident angle was studied in other cases (Dalila's PhD).

Could we use Geant4 for the early stages of the shower, then break out to ML once energy falls below a threshold. This would then be a much easier training task.

Data format conversions - improvements coming soon on ROOT side (worth speaking to Sitong).

There are minutes attached to this event. Show them.