Accelerating the Search for Dark Matter with Machine Learning

15 Jan 2018, 09:00 → 19 Jan 2018, 14:45 Europe/Zurich

Lorentz Center@Oort

Organisers: Tom Heskes (Radboud University), Gianfranco Bertone (UvA), Francesca Calore (CNRS), Sascha Caron (Radboud University and Nikhef) and Roberto Ruiz de Austri (IFIC Valencia)

Contact: Nieke Tander (Lorentz Center), Sascha Caron

From Monday 15 January through Friday 19 January 2018

Lorentz Center@ Oort – Leiden, The Netherlands
 

Scientific Topic : In this workshop we aim to explore, and to encourage, the utilization of state-of-the-art machine learning algorithms for research in dark matter physics and astronomy. Our objective is to accelerate the identification of Dark Matter with a multidisciplinary approach: The researchers coming to this workshop bring together expertise in experimental and theoretical particle physics, astrophysics, astronomy, statistics and Machine Learning. The workshop is planned to be a Kickoff meeting to generate a new open research community. We plan a whitepaper, follow-up workshops, a webpage and a mailing list for the DM-ML community.

Workshop Format Lorentz Workshops@ Oort are scientific meetings for small groups of up to 60 participants, including both senior and junior scientists. Lorentz Center meetings dedicate a considerable amount of time to discussion sessions, thus stimulating an interactive atmosphere and encouraging collaborations between participants. This format typically generates extensive debates and enables significant progress to be made within the research topic of the meeting.

Registration by 15 October via:

Link to the Workshop webpage at the Lorentz Center

Lorentz Center Facilities The venue Lorentz Center@ Oort, located at the Faculty of Science campus of Leiden University, the Netherlands. www.lorentzcenter.nl.

 

Monday, 15 January

1 Arrival / Registration

2 Welcome from Lorentz Center

3 Welcome, Workshop Structure and Objectives

Speaker: Sascha Caron (Nikhef National institute for subatomic physics (NL))

DMMLworkshop_intro.pdf

4 Dark Matter overview

Speaker: Nicolao Fornengo (University of Torino and INFN)

fornengo.pdf

5 Introduction into Deep Learning and Image Analysis

Speaker: Max Welling

LorentzWorkshopDarkMatter2018.pdf

LorentzWorkshopDarkMatter2018.pptx

6 First ideas to connect Astronomical data, Deep Learning and Image Analysis

Speakers: German Arturo Gomez Vargas (Universidad Autonoma de Madrid), German Gomez-Vargas (Pontifical Catholic University of Chile)

first_ideas_astronomy_ML_Gomez-Vargas_Leiden_2018.pdf

7 Discussion: First connections ?

12:35 Lunch

8 Cross link: Deep Learning in High Energy Physics

Speaker: Amir Farbin (University of Texas at Arlington (US))

MLDM.pdf

9 Discussion: New ideas ?

Bring up projects --> Bring up a discussion ...

We can vote which one we should open.

Decide on the projects (darksurvey.com)

Suggest new application of Deep Learning / Image Analysis.
Suggest topic for brainstorming session.

15:05 Coffee break

10 Work in subgroups : Finding new projects in Deep Learning / Image Analysis

Person proposing the subject + Organiser + Participants
split up in 2-4 rooms.

Work on 5 min presentation about the topic.

11 The hunt for stellar-mass DM clumps: applying the statistical machine learning techniques to strong microlensing events

Strong gravitational microlensing (GM) events give us a possibility to determine some characteristics of both microlens and microlensed source. As the role of microlens can be played by a DM clump, GM can give us an important clue to understand the nature of dark matter on comparably small spatial/mass scales. In the same time, fitting the lightcurves of microlensed sources is quite time-consuming process, especially taking into account nonzero lens size. Here we test the possibility to apply the statistical machine learning techniques to distinguish high-amplification microlensing events (HAME) caused by continuously distributed DM clump from star- or black hole- induced microlensing (i.e. microlens is considered as a point-like mass). On this stage we use the set of simulated HAE amplification curves of sources microlensed by point masses and clumps of DM with various density profiles.

Speaker: Elena Fedorova

The hunt for stellar-mass DM clumps.pdf

12 Fast model discrimination with Euclideanized signals

Speaker: Christoph Weniger (University of Amsterdam)

ML4DM.pdf

17:00 Wine and Cheese Party at Common Room (maybe Posters)

Tuesday, 16 January

13 Results of yesterdays workgroups on Deep Learning, Image Analysis and astronomical Dark Matter data

14 Astronomical Dark Matter measurements and Challenges

Speaker: David Richard Harvey (EPFL - EPF Lausanne)

Leiden_18.pdf

10:15 Coffee break

15 Introduction into direct and indirect Dark Matter searches and their challenges

Speakers: Dr Marco Regis (INFN - National Institute for Nuclear Physics), Marco Regis

DM&ML_Regis.pdf

16 Introduction into unsupervised learning

Speaker: Erzsébet Merényi

UnsupervisedLearning-Merenyi.pdf

17 Discussion Unsupervised Leaning for Dark Matter

12:25 Lunch

18 First ideas to use Machine Learning in direct Dark Matter searches

Speakers: Andrew Brown (MIT), Andrew Brown (Nikhef), Christopher Tunnell (Enrico Fermi Institute-University of Chicago-Unknown), Dr Christopher Tunnell (University of Chicago)

19 Discussion: New ideas ?

20 First ideas to use Machine Learning in indirect detection

Speaker: Luc Hendriks (Nikhef)

Presentation_LHendriks.pdf

21 Discussion: New ideas?

15:30 Coffee

22 Work in subgroups

23 Characterization of the Local Universe via angular cross-correlations

Speaker: Simone Ammazzalorso (University of Turin)

Leiden_2018_indico.pdf

24 Dark matter searches in dwarf irregular galaxies

Speaker: Viviana Gammaldi (SISSA)

Gammaldi_Leiden_v1.pdf

Wednesday, 17 January

25 Results of yesterdays workgroups on unsupervised learning, direct and indirect DM searches

26 New approaches in semi-supervised learning

Speaker: Kristiaan Pelckmans

presentation_semisupervised.pdf

10:30 Coffee

27 First ideas: Supervised DNNs for reconstruction in LHC data

Speaker: Markus Stoye (CERN)

datascience.pdf

28 Recent results : Dark Matter searches at LHC

Speakers: Renjie Wang (LPNHE-Paris CNRS/IN2P3 (FR)), Renjie Wang (LPNHE-Paris, CNRS/IN2P3 (FR))

DM_LHC.pdf

29 Machine Learning for SHiP and NEWS experiments

Emulsion-based detectors such as ones used for OPERA experiment or planned for SHiP and NEWS experiments may reveal important characteristics of WIMP-like particles. However due to the nature of the emulsion, the signal to noise ratio tend to be rather small and hence might require special reconstruction techniques. Thus advanced data analysis approaches based on machine learning approaches might improve «physical» sensitivity of the experiments. In this talk I’ll give brief overview of machine learning techniques that can be applied for dark matter searches in SHiP and NEWS experiments and present current challenges for those experiments both from physical and data analysis points of view.

Speaker: Andrey Ustyuzhanin (Yandex School of Data Analysis (RU))

Ustyuzhanin-dark-yandex-v2.pdf

12:30 Lunch

30 Miscellaneous thoughts on Machine Learning & Dark Matter

Speaker: Kyle Stuart Cranmer (New York University (US))

ML-for-Dark-Matter.pdf

31 Discussion: New ideas? Reinforcement Learning?

15:05 Coffee

32 Work on projects in subgroups

33 Data-driven constraints on dark matter from dwarf galaxies

Speaker: Mr Bryan Zaldivar (LAPTh, Annecy)

zaldivar_Leiden2018.pdf

34 Fast Forecasting for Counting Experiments

Speaker: Tom Edwards

Presentation_TEdwards.pdf

35 Estimating the parameters of gravitational lenses with deep learning

Machine learning methods have seen a rapid expansion in the last few years. In particular, deep learning has made several breakthroughs, including beating a champion of game of Go and outperforming practicing dermatologists in the visual diagnosis of skin cancer. Although in most applications these networks have been used for classification tasks, they can also be made to predict real-valued model parameters. In this talk, I will discuss our results on using deep convolutional neural networks to estimate the parameters of strong gravitational lenses from telescope data. Estimating these parameters with traditional maximum-likelihood modeling methods is a time and resource consuming procedure, involving several data preparation steps and a difficult optimization process. With deep convolutional neural networks we are able to estimate these parameters in a fully automated way 10 million times faster than traditional modeling methods and with a similar accuracy. I will also discuss how to robustly quantify the uncertainties of these networks. This allows them to be a fast alternative to MCMC sampling. With the advent of large volumes of data from upcoming ground and space surveys and the remarkable speed offered by these networks, deep learning promises to become an indispensable tool for the analysis of large survey data.

Speaker: laurence perreault levasseur (Stanford University)

Leiden_PerreaultLevasseur.pdf

17:00 Conference Dinner (Boat trip)

Thursday, 18 January

36 Results of yesterdays workgroups on supervised learning and DM searches at the LHC

37 Introduction into theory models for Dark Matter particles

Speaker: Andrea de Simone

DeSimone_DMmodels.pdf

10:15 Coffee

38 OpenML/AutoML: Organizing machine learning data and learning to learn better models

Speaker: Joaquin Vanschoren

OpenML Lorentz.pdf

39 Adversarial Games for Particle Physics

Speaker: Gilles Louppe (New York University (US))

slides.pdf

40 Discussion: Further ideas?

12:25 Lunch

41 Dark Matter model exploration and first ML ideas

Speaker: Martin John White (University of Adelaide (AU))

Leiden.pdf

42 How to find Natural Supersymmetric Dark Matter?

Supersymmetry (SUSY) is able to solve the hierarchy problem and it can provide a perfect dark matter candidate. The non-observation of SUSY particles at the LHC and dark matter particles at dedicated experiments drives the SUSY particles to be heavier and heavier, which is assumed to make it more and more difficult for SUSY to solve the hierarchy problem as it gives rise to the need of fine-tuning of the input parameters of the theory. We are studying the allowed parameter space of several SUSY models. These models typically have a large number of parameters (10-30). We aim to find the set of allowed parameters that minimize the fine-tuning of these SUSY models. This is a resource-consuming process and we would like to discuss on how to do this more efficiently.

Speaker: Melissa van Beekveld (R)

43 BSM-AI (SUSY-AI) and iDarkSurvey: Learning (from) high-dimensional models

Although the standard model of particle physics is successful in describing physics as we know it, it is known to be incomplete. Many models have been developed to extend the standard model, none of which have been experimentally verified. One of the main hurdles in this effort is the dimensionality of these models, yielding problems in analysing, visualising and communicating results. Because of this, most current day analyses are done using simplified models, but in this process descriptive power is lost. However, by using machine learning on simulated model points, we show that we can overcome these problems and predict both binary exclusion an continuous likelihood in any parameter space. This simulated data will be stored in our new webbased database and model visualisation tool iDarkSurvey. This tool will be open to the scientific to store all calculated model data.

Speaker: Bob Stienen (Radboud University)

44 Using Deep Learning to predict Electroweakino production cross-sections at the LHC

Speaker: Sydney Otten (RWTH Aachen)

15:25 Coffee

45 Discussion: New ideas?

46 Work in subgroups

Friday, 19 January

47 Open discussion: what have we leart? What next?

11:30 Coffee and Goodbye

12:00 Lunch