Optimisation of the SHiP muon shield Oliver Lantwin on behalf of the - PowerPoint PPT Presentation

Optimisation of the SHiP muon shield Oliver Lantwin on behalf of the SHiP Collaboration. IoP app/hepp March 26, 2018 [ oliver.lantwin@cern.ch ]

The current state of physics “We know there is new physics,…” Dark matuer, baryon asymmetry and neutrino masses are direct experimental evidence that we’re missing something. “… We don’t know where it is…” We do not know which energy scale to target: Very weakly coupled new physics could be hiding in plain sight — at energies already accessible! “… We need to be as broad as possible in our exploratory approach” Oliver Lantwin (Imperial College London) IoP app/hepp Introduction 2 — Fabiola Gianotui

Overview of the Search for Hidden Particles Target & Magnetised hadron absorber The SHiP Experiment IoP app/hepp Oliver Lantwin (Imperial College London) new physics models Generic signatures predicted by many Zero Background crucial to study hidden sector decays 2. Via scatuering in nuclear emulsion 1. Via decay to visible particles in hidden sector spectrometer Two signatures: 3 𝜈 𝜌 Yields for 2 × 10 20 pot (5 years): Hidden sector spectrometer Decay volume Emulsion spectrometer Active muon shield > 10 18 𝐸 , > 10 16 𝜐, but 10 18 𝜈 hnl 𝑞 @400 GeV m 5 1 1

Crucial challenge: Zero background available in simulation The SHiP Experiment IoP app/hepp Oliver Lantwin (Imperial College London) ⨂ ⨀ SHiP 𝑨[ m ] 𝑧[ m ] 4 › Passive hadron absorber for this summer target at the h4 test-beam at cern’s sps is planned › A measurement of the muon spectrum for the SHiP maximise the experimental acceptance The muon shield is the critical component to optimise to › kinematic range of muons up to 𝑞 ∼ 350 GeV at least 6 orders of magnitude [2017 JINST 12 P05011] › Active muon shield that has to reduce muon flux by › kinematic range of muons up to 𝑞 𝑈 ∼ 8 GeV › Obtain 10 11 protons on target, c.f. 10 10 currently

Goals & Challenges of the muon shield optimisation → Evaluation of points very expensive, gradient information not available and can not be Muon shield optimisation IoP app/hepp Oliver Lantwin (Imperial College London) from cm to m › Even with a simple parametrisation we have ~50 free parameters (lengths), each varying approximated › With a difgerent random seed entirely difgerent muons pass the shield › Nearly identical configurations may have very difgerent performance › Underlying physics inherently stochastic › Not enough computing power to use entire simulation for optimisation › Not enough simulation › Doubly statistically limited Challenges optimise performance vs. cost and provide robustness by optimising for a lower field strength . 5 Goal: Optimisation using full simulation with FairShip framework for every evaluation to

Introduction to Bayesian Optimisation using a 1D example * Oliver Lantwin (Imperial College London) Muon shield optimisation IoP app/hepp 6 x + = 0 . 1000 t 1.5 True (unknown) Observations µ GP ( x ) 1.0 u(x) CI 0.5 f(x) 0.0 0.5 1.0 1.5 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 x * Based on scikit-optimize documentation

How we use Bayesian Optimisation space: Muon shield optimisation IoP app/hepp Oliver Lantwin (Imperial College London) and computing power † Russian internet company which contributes to lhc b , comet, cms and SHiP with its machine learning expertise › Evaluating importance sampling and other options 2. reduce and re-weight manually 1. study the importance of difgerent regions of the phase-space › Currently: 7 Not quite as simple as this example: › Use Gaussian processes and random forests as surrogate models. › Make up to 100 guesses at once (with 16 nodes parallelising every function evaluation) › 1600 cores available at Yandex † › Computing model imposes additional constraints. › Bayesian optimisation does not scale well for high-dimensional problems. › Use scikit-optimize implementation of Bayesian optimisation DOI 10.5281/zenodo.1170575 10.5281/zenodo.1170575 . DOI › Reduce muon sample by factor ~ 40 to speed up evaluation and even out coverage of phase

Loss function sensitive plane at position 𝑦 𝜈 . Muon shield optimisation IoP app/hepp Oliver Lantwin (Imperial College London) background studies. Loss function continues to evolve with technological constraints and Figure 1: 𝜓 𝜈 (𝑦 𝜈 ) 𝑔 (𝑋, 𝜓 𝜈 ) = › Weight cut-ofg as regularisation › Length optimised implicitly via the weight › Penalise muons entering the acceptance Note: 8 ⎨ 10 8 𝑋 weight of the muon shield where: otherwise, ⎧ { if 𝑋 > 3 kt { ⎩ (1 + exp (10 × (𝑋 − 𝑋 0 )/𝑋 0 )) × (1 + ∑ 𝜈 𝜓 𝜈 (𝑦 𝜈 )) µ + 1 µ − 0 . 8 𝑋 0 weight of the baseline 0 . 6 χ µ 𝜓 𝜈 weighted position of muon 𝜈 passing a 0 . 4 0 . 2 0 − 3 − 2 − 1 0 1 2 3 x µ / m

Optimisation convergence reduced muon sample: perform Muon shield optimisation IoP app/hepp Oliver Lantwin (Imperial College London) › Cumulative loss: exploring dataset to confirm performance follow-up studies on the full 9 › Performance here is on the algorithms to determine which evaluating difgerent regression › Two optimisers shown here: still points with high uncertainty part of algorithm, only cumulative loss is meaningful performs best loss function 10 1 cumulative minimum loss rf cumulative minimum loss gb baseline 10 0 0 . 0 0 . 5 1 . 0 1 . 5 2 . 0 2 . 5 × 10 3 iteration

Results 70±15 Muon shield optimisation IoP app/hepp Oliver Lantwin (Imperial College London) 42±6 22±3 1.28 34.82 new optimum @1.7 T 27±5 › Significant reduction in weight (→cost) 1.72 34.60 baseline @1.8 T full sample reduced sample weight/kt length/m Configuration › Same performance with significantly reduced magnetic field 10

Technology & Prototyping › Several prototypes will be produced this year, and Muon shield optimisation IoP app/hepp Oliver Lantwin (Imperial College London) technology cern → Part of the cern/Imperial team testing the the most promising will be tested with beams at Optimise technology as well as geometry Grain oriented steel joints of the magnets › Several techniques need to be evaluated for the › Scale of muon shield exceptional technology: › Manufacturing of SHiP will push the limits of the magnets › Allows to achieve fields of up to 1.8 T with warm 11

Conclusion and further work › Found new configuration for comprehensive design study. › Have an algorithm that works and can be used as base for further improvements. › Optimisation infrastructure is now also used for optimisation of other subsystems. Future work › Fully automate process, add additional constraints to loss function and improve the shield further! › Collaboration with engineers at misis to progress to a detailed engineering design and prototypes. Oliver Lantwin (Imperial College London) IoP app/hepp Conclusion 12

Backup Oliver Lantwin (Imperial College London) IoP app/hepp Backup 13

Crucial challenges Maximise intensity and mass reach › Intense proton beam from the sps @400 GeV at the new beam dump facility (bdf) in the North Area › Very dense target of 12 × 𝜇 int › abundant production of heavy flavour › Number of protons per cycle similar to cngs, but slow instead of fast extraction › Operation in parallel with lhc, other beam-lines at the sps Oliver Lantwin (Imperial College London) IoP app/hepp Backup 14 › reduced neutrino production from 𝜌 and 𝐿 decays

Sensitivity: hnl › Baryon asymmetry of the universe (bau) Backup IoP app/hepp Oliver Lantwin (Imperial College London) NB: Before re-optimisation model › Model-independent limit for any Seesaw › Big bang nucleosynthesis (bbn) Theoretical limits from: Figure 2: hnl sensitivity at SHiP for 𝜉 msm with › Significant contribution from 𝐶 -decays limit › Best sensitivity up to charm kinematic neutrino mass hierarchy. 𝑉 2 15 𝑓 ∶ 𝑉 2 𝜈 ∶ 𝑉 2 𝜐 = 1 ∶ 16 ∶ 3.8 and a normal

Sensitivity: Dark Scalars Figure 3: Dark scalar sensitivity at SHiP. › For short lifetimes 𝐶 -factories and LHCb best › SHiP covers unique parameter space complementing other experiments › Large contribution from 𝐶 -decays at SHiP › “Hole” at 𝑑𝜐 ∼ 𝒫( m ) , where lifetime is too short for SHiP and too long for 𝐶 -experiments NB: Before re-optimisation Oliver Lantwin (Imperial College London) IoP app/hepp Backup 16

Sensitivity: Dark Photons by other experiments Backup IoP app/hepp Oliver Lantwin (Imperial College London) NB: Before re-optimisation determined by short lifetime › Top-right edge of sensitivity › Complementary to regions studied Figure 4: Dark photon sensitivity at SHiP. → Work in progress and meson decays › Produced in qcd, bremsstrahlung › Visible decays of dark photons years › Based on > 10 20 𝛿 at SHiP over 5 17 › No production via em showers yet