BOB0 : Bayesian Optimization for B0

Description

This repository provides a Bayesian optimization module for B0, a far-forward detector sub-system of ePIC [1], the future detector setup on EIC (Electron-Ion Collider) [2]. The optimized parameters are the positions along the z-axis of the four tracker disks.

The optimization is computationnaly intensive: the module is designed for running on iFarm with Slurm. It also includes a checkpoint manager to start and stop at any time. The checkpoint files (to clean if you want to start from scratch) are saved in checkpoints/.

Model & inputs

The Bayesian optimization [3], based on the BoTorch library [4], is based three functions:

• Surrogate function (SF): fixed as Gaussian processes (GP).
• Acquisition function (AF): fixed as Noisy Expected Improvment (logNEI).
• Objective function (OF) : to be chosen by the user (see section Use).

The inputs, to be chosen by the user, are encoded in .hepevt files. They are stored in /events, and include:

• Proton gun at 41 GeV (200k) along z-axis with 25 mrad tilting.
• Proton gun at 100 GeV (200k) along z-axis with 25 mrad tilting.
• Proton gun at 275 GeV (200k) along z-axis with 25 mrad tilting.
• $\Lambda$ from Sullivan process at 5 x 41 GeV (500k).
• $\Lambda$ from Sullivan process at 10 x 100 GeV (500k).
• $\Lambda$ from Sullivan process at 18 x 275 GeV (500k).

Remark: BOB0 simulations based on Geant4 manage instable particles and decay products, especially for $\Lambda$.

HPC strategy

BOB0 divides the input file into T x C chunks, where T is the number of tasks and C the number of CPUs per task. Both T and C can be customize by the user (see section Use).

A volley of T small jobs are launched, each one managed by C CPUs, and saved in bayesian_iteration/. Simulation results from all jobs are finally merged into a unique file (.root) saved in simu/. This file will by analyzed to evaluate the objective function.

Installation

Installing BOB0 is a 5-steps process.

Step 1. Clone the repo on your personal working space in iFarm. It will create the project bob0/.

git clone git@github.com:baptistefraisse/bob0.git

Step 2. Install eic-shell environment (the EIC group container) within the project.

cd bob0
mkdir eic
curl -L https://github.com/eic/eic-shell/raw/main/install.sh | bash

Step 3. Install an old version of eic-shell compatible with EICrecon on the EIC branch we'll work with. This step requires the singularity package (available on iFarm). Creation of the .sif file may be a long step (few hours). Once this file has been created, change the very last line of .sif with your project path.

cd bob0
mkdir Old_eic-shell
mkdir $PWD/tmp && export SINGULARITY_CACHEDIR=$PWD
singularity pull --tmpdir=$PWD/tmp --name eic-jug_xl-nightly-2024-03-12.sif docker://eicweb/jug_xl@sha256:213e55fb304a92eb5925130cdff9529ea55c570b21ded9ec24471aa9c61219d8
cp ../eic/eic-shell ./custom-eic-shell

Step 4. Install the ePIC model for Geant4 simulations inside the custom-eic-shell environement. We'll work with a branch including a realistic magnetic field inside B0 called b0-field-map-testing. After cloning and before installing, comments the lines 27 to 29 in CMakeLists.txt (stop at any warning).

cd bob0
cd Old_eic-shell
source custom-eic-shell
git clone https://github.com/eic/epic.git -b b0-field-map-testing
cd epic # Here comment lines 27-28-29 in CMakeLists.txt
mkdir epic_install epic_build
cmake -B ./epic_build -S . -DCMAKE_INSTALL_PREFIX=./epic_install
cmake --build ./epic_build -j8
cmake --install ./epic_build

Step 5. Install EICrecon. We'll checkout the version v1.11.0 compatible with our EIC environement (custom-eic-shell) and the b0-field-map-testing branch of ePIC.

cd bob0
cd Old_eic-shell
source custom-eic-shell
source ./epic_install/setup.sh
git clone https://github.com/eic/EICrecon.git
cd EICrecon
git checkout v1.11.0
mkdir EICrecon_build EICrecon_install
cmake -B ./EICrecon_build -S . -DCMAKE_INSTALL_PREFIX=./EICrecon_install
cmake --build ./EICrecon_build -j8
cmake --install ./EICrecon_build

Use

Launching BOB0 is a 3-steps process. The user interface lay in the bobo.env file.

Step 1. In bobo.env, change the project direction afer the dash. For example, mine is:

export PROJECT_DIR="${PROJECT_DIR:-/volatile/eic/fraisse/bob0}"

Step 2. In bobo.env, customize the Bayesian optimization parameters: parallelization parameters, bounds, number of initialization and iterations, objective function and inputs. By default:

export NARRAYS=100 # T
export NCPUS_PER_TASK=8 # C
export LOWER_BOUNDS='[0, 10, 10, 10]' # z1, dz2, dz3, dz4
export UPPER_BOUNDS='[40, 40, 40, 40]' # z1, dz2, dz3, dz4
export N_INIT=20 # number of init iterations
export N_ITER=50 # total number of iterations
export OBS_OBJ='p' # total momentum
export EVENTS_PREFIX=gun_proton_100GeV # 100 GeV protons

Step 3. Launch BOB0 with Slurm using sbatch command with job.mobo.slurm.sh. Status can be checked using squeue. Jobs can be killed using scancel.

sbatch job.mobo.slurm.sh

Outputs

The progress of BOB0 can be followed through two files, .out and .err, written in dedicated folder logs/. The final result (optimized z-positions) will be written at the very end of .out. The figures, including the convergence function, are saved in the dedicated folder fig/.

Outlooks

Some improvements are planned or ongoing, including:

• [computational] Upgrade to multi-objective optimization (Paretto frontier).
• [computational] Upgrade to asynchronous Bayesian optimization (q > 1) [5].
• [physics] Optimization of active/dead zones of disks surfaces (Zernike decomposition).
• [physics] Holistic optimization of a physics channel (Sullivan process [6]).
• [cluster] Move from Slurm to Swif2.
• [cluster] Use of GPUs.

References

[1] M. Pitt, Proc. of DIS2024, arXiv:2409.02811 (2024)

[2] Yellow Report on EIC, Nucl. Phys. A 1026, 122447 (2022)

[3] B. Shahriari et al., Proc. of the IEEE (2016)

[4] M. Balandat et al., Advances in Neural Information Processing Systems (2020)

[5] B. Shahriari et al., Proc. of the IEEE (2016)

[6] J. D. Sullivan, Phys. Rev. D 5, 1732 (1972)

Contact

fraisse@cua.edu