Physics Applications of AI Day

Wednesday 3 Nov 2021, 13:30 → 18:00 Europe/Zurich

Ecole de physique

The Physics Applications of AI Day is organised by the Physics Section of the Faculty of Science and it aims at:

• Presenting the research activities of the Physics, Astronomy and Computer Science Departments in the domain of Artificial Intelligence (AI) and Machine Learning (ML)
• Interaction of junior researchers and in particular PhD students among the Departments
• Fostering collaboration between the Departments of Physics, Astronomy and Computer Science of the Faculty of Science
• Kicking off a new master-level course given in the Spring semester 2022 and open to bachelor students as well PhD students (link)

The event will take place on Wednesday, November 3rd, 2021 at 13:30 in the Ecole de Physique, Grand Auditoire, starting with overview and invited talks followed by a poster session with deluxe sandwiches (Boucherie du Palais/Traiteur Vidonne) and drinks in the halls of the building.  A covid certificate is required for attending the poster session.

The event is open to the wider Faculty of Science community as well as all master and bachelor students.

The talks will be recorded and there is a livestream: https://mediaserver.unige.ch/live

---

Please register for the Physics Applications of AI Day using the registration tab.  Registration deadline is set to October 26 2021.

Once you have registered please submit your poster title through the abstract submission.

---

Scheduled Talks

 

	Jerome Lacour - Welcome
	Tobias Golling - AI & DPNC
	Slava Voloshynovskiy - Information-theoretic imaging and classification
	Patrycja Paruch - AI & DQMP
	Nathan Hara - AI & Astro
	Francois Fleuret - Machine learning and deep learning based simulation proxies
	Jean-Pierre Wolf - AI & GAP
	Benjamin Bose - AI & Theory

13:30 → 16:00 Overview Talks

Convener: Prof. Tobias Golling (University of Geneva)

13:30 Welcome

Speaker: Prof. Jérôme Lacour

13:32 Welcome

Speaker: Tobias Golling (University of Geneva)

13:40 AI & DPNC

Speaker: Prof. Tobias Golling (University of Geneva)

14:00 Information-theoretic imaging and classification

Speaker: Slava Voloshynovskiy (UniGE)

14:20 AI & DQMP

Speaker: Patrycja Paruch

14:40 AI & Astro

Speaker: Nathan Hara

15:00 Machine learning and deep learning based simulation proxies

Speaker: François Fleuret (University of Geneva)

15:20 AI & GAP

Speaker: Jean-Pierre Wolf

15:40 AI & Theory

Speaker: Benjamin Bose

16:00 → 18:00 Poster & Deluxe Sandwich Session

16:00 ML in ATLAS

overview poster of ML in ATLAS and showing some topics for which students could join the group

Speakers: Manuel Guth (Université de Genève) , Johnny Raine (Universite de Geneve (CH))

16:01 Generation of data on discontinuous manifolds via continuous stochastic non-invertible networks

The generation of discontinuous distributions is a difficult task for most known frameworks, such as generative autoencoders and generative adversarial networks. Generative non-invertible models are unable to accurately generate such distributions, require long training and often are subject to mode collapse. Variational autoencoders (VAEs), which are based on the idea of keeping the latent space to be Gaussian for the sake of a simple sampling, allow an accurate reconstruction, while they experience significant limitations at generation level. In this work, instead of trying to keep the latent space Gaussian, we use a pretrained contrastive encoder to obtain a clustered latent space. Then, for each cluster, representing a unimodal submanifold, we train a dedicated low complexity network to generate it from the Gaussian distribution. The proposed framework is based on the information-theoretic formulation of mutual information maximization between the input data and latent space representation. We derive a link between the cost functions and the information-theoretic formulation. We apply our approach to synthetic 2D distributions to demonstrate both reconstruction and generation of discontinuous distributions using continuous stochastic networks.

Speaker: Mariia Drozdova (Universite de Geneve (CH))

16:02 Transformers and Graph Neural Networks

-

Speaker: Daniele Paliotta (UniGe)

16:03 Deep learning applied to X-ray tomography as a new tool to analyse internal properties of superconductive wires

Tommaso Bagni¹, Diego Mauro¹, Gianmarco Bovone¹, Marta Majkut², Alexander Rack², Carmine Senatore^1
1Department of Quantum Matter Physics, University of Geneva, Geneva, Switzerland
²ESRF – The European Synchrotron, Grenoble, France

Speaker: Tommaso Bagni (University of Geneva)

Poster_PhysApplAI_TBagni_ver0.3.pdf

16:04 Variable object classification for the third Gaia Data Release (DR3)

We presents a short summary of the variable object classification effort for the third Gaia Data Release (DR3) that will become public mid 2022. Some example figures of the previous (public) DR2 release are shown to illustrate our validation process.

Speaker: Berry Holl (University of Geneva, Geneva Observatory)

VariableObjectClassificationForGaiaDataRelease3.pdf

16:05 Neural networks for electron identification with DAMPE

The past decades have witnessed the deployment of a new generation of cosmic ray (CR) observatories with unprecedented sensitivity and complexity, pushing towards ever-higher energies. To face the challenges of the multi-TeV domain, such instruments must be accompanied by equally powerful analysis techniques, able to exploit the wealth of available information. The machine learning tool set may provide the needed techniques. I present a neural network optimised for the identification of multi-TeV electrons on DAMPE, a calorimetric spaceborne CR observatory with among other objectives the measurement of cosmic electrons up to 10 TeV. This constitutes a particularly challenging endeavour due to both the soft electron spectrum and the large proton background. The developed neural network significantly outperforms the more traditional cut-based approach, achieving a much lower proton contamination in the multi-TeV domain with a high signal efficiency, and retains its accuracy when transposed from Monte Carlo to real data.

Speaker: David Droz (University of Geneva)

poster_PHML_DPNC.pdf

16:06 Solving ttbar Combinatorics Using Graph Neural Networks

Events with a $t\bar{t}$ pair can have up to six jets coming from the two top quarks. Assigning these jets correctly to the two quarks is challenging due to large combinatorics especially in the allhadronic final state. The correct assignment of these jets has several benefits. For instance, the kinematics of the top quarks can be determined. Several different methods already exist for reconstructing different ttbar final states, both with and without machine learning. The novelty of this approach is the insertion of helper nodes as the intermediate particles. These helper nodes can additionally be used to regress towards the true properties of the intermediate particles. The approach can be generalized to any sort of decay chain within a detector

Speaker: Lukas Ehrke

ttbarCombinatoricsPhysicsAIPoster.pdf

16:07 Jet Flavour tagging

Jet flavour tagging using machine learning (DIPS, Umami)

Speakers: Tomke Schröer, Dimitri Moulin (Université de Genève)

Jet_Flavour_tagging.pdf

16:08 Novel functionalities at twin domain crossings

In ferroelectrics, domain walls are thin interfaces separating regions with different orientations of electric polarization, either along the same crystalline axis (180° domain walls), or as ferroelastic twins. The domain walls can present physical properties quite different from the surrounding domains, allowing them to be used as active components in future device applications.

Recent studies of domain walls using scanning probe microscopy have focused on mapping their response to different parameters such as temperature, applied pressure and electric field, in order to understand their structure-property relationships. In particular, the role of high strain gradients present at ferroelectric twins has been shown to enhance their electrical conduction [1] and can lead to complex rotational polarization textures [2,3].

Here, I will present our investigation of ferroelastic twin domains (90° domain walls) in epitaxial PbTiO3 thin films grown on SrTiO3, explored with scanning probe microscopy. Our results suggest a complex polarization structure, with unique mechanical response distinct from the surrounding ferroelectric phase, and enhanced electrical conduction.

References
[1] Stolichnov, I., et al. Nano Lett. 15, 8049−8055 (2015)
[2] Catalan, G., et al, Nat. Mat. 10, 963 (2011)
[3] Cao, Y., et al, Appl. Phys. Lett. 110, 202903 (2017)

Speaker: Kumara Cordero-Edwards (DQMP, University of Geneva)

16:09 Hystorian: A Processing Tool for Scanning Probe Microscopy and Other n-Dimensional Datasets

In recent years, data processing has held an increasingly important role in the toolkit of a scanning probe microscopist. From simple parabolic baseline correction in a topography image, to utilising correlations between images to adjust for distortions between scans, external programming platforms have allowed the extraction of information well beyond the raw data.

This information revolution has also yielded new challenges for the modern researcher. Complex processing protocols require conversion from proprietary formats into user-defined alternatives. These alternatives may lose key physical metadata in the process, such as scale, scan rate, or applied tip voltage. Furthermore, as data-processing programs are often written by individual users or lab groups, potential deviations can appear when comparing results with other researchers. Such issues are exacerbated if the user attempts to collate data from independent sources, where a process may have to be completely rewritten to operate on another data type.

Here, we present Hystorian, a cross-platform Python library that is capable of loading, merging, and operating on data from arbitrary sources [1], by transforming the data into n-dimensional arrays in a hierarchal dataset format (HDF5) file. All metadata is stored in raw text as well as in attributes associated with the primary dataset. Processing functions can therefore utilise multiple datasets in a single datafile, in addition to their associated metadata. The history of process outputs is also saved within the file, allowing a user to operate on intermediate processes, trace a history of these operations, and access all outputs without reapplying computationally intensive algorithms.

Hystorian is also packaged with tools to assist in the application of custom functions to the HDF5 files, along with a range of useful scientific operations. This allows for rapid prototyping and accessibility to any user with a working understanding of Python. The source-agnostic nature of the HDF5 file also allows similar functions to be used across different data types. For example, feature identification in scanning probe microscopy images may be directly transferred to track changes in peak positions in x-ray diffraction reciprocal space maps. In this poster, Hystorian is used to process a series of 30 piezoresponse force microscopy images to track the motion of domain walls over a 20-hour period.

[1] Musy, L., Bulanadi, R., et al. "Hystorian: A processing tool for scanning probe microscopy and other n-dimensional datasets." Ultramicroscopy 228 (2021): 113345.

Speakers: Loïc Musy (University of Geneva) , Ralph Bulanadi (University of Geneva)

16:10 Finding gravitational wave signals from binary black hole collisions with convolutional neural networks

In 2015, the first gravitational wave signals from colliding binary black holes were detected. Subsequent detections of gravitational waves lead researchers to observe a new population of massive, stellar-origin black holes. These signals are tiny ripples of the fabric of space-time. Even though the global network of gravitational-wave detectors is one of the most sensitive instruments on the planet, the signals are buried in detector noise. Analysis of gravitational-waves data and the detection of these signals is a crucial mission for the gravitational-waves community. In this project, I used a machine learning technique to analyse simulated gravitational-wave time-series data from a network of ground-based detectors. More precisely, time-series data from three detectors were converted into spectrograms using a constant Q-transform and combined into a single RGB figure. Subsequently, I used transfer learning to train a state-of-the-art convolutional neural network to detect gravitational wave signals from the merger of binary black holes.

Speaker: Simone Bavera (University of Geneva)

SB_poster.pdf

email: simone.bavera@ungie.ch

16:11 Full LST-1 data reconstruction with the use of convolutional neural networks

The Cherenkov Telescope Array (CTA) will be the world's largest and the most sensitive ground-based gamma-ray observatory in the energy range from a few tens of GeV to tens of TeV. To achieve the best performance in such a wide range of primary gamma-ray photon energies, the whole array will consist of Imaging Atmospheric Cherenkov Telescopes of three different sizes. The Large-Sized Telescopes (LSTs) are designed to be the most sensitive in the low-energy band of CTA. The LST-1 prototype, currently in the commissioning phase, was inaugurated in October 2018 in La Palma (Spain) and it is the first of the four largest CTA telescopes, that will be built in the northern site of CTA.

The Random Forest method is currently employed in the reconstruction of the first data from the LST-1. This method, however, requires extensive preprocessing that includes signal integration and image parametrization, leading to the loss of information contained in the shower images. In this contribution, we will present a full-image reconstruction method using Inception deep convolutional neural network (CNN) applied on non-parametrized shower images. First, we will evaluate the performance of optimized networks on Monte Carlo simulations of LST-1 shower images, and compare the results with the performance of the standard reconstruction method. We will also show how both methods compare to real-data reconstruction.

Speakers: Jakub Jurysek (University of Geneva - Department of Astronomy) , Etienne Lyard (University of Geneva - Department of Astronomy)

16:12 AI-based downy and powdery mildew detection in microscopy and holography in vines

The measurement of pathogens in field samples with in-lab microscopy produces massive amount of images (spores $\sim 5 – 30 \mu$m and scanned surface $\sim 3 $cm$^2$) whose analysis requires some automation. A trained AI-model of instance segmentation aims to recognize spores within images containing a multitude of other particles of similar sizes and/or displaying analogous textures.

Conversely, a device has been developed to capture airborne particles in the field and, based on their distinctive holographic patterns, the presence of spores of interest is estimated through an AI-model for object detection.

Speakers: Mrs Tessa Basso (GAP) , Michael Tran

16:13 Priming PCA with EigenGame

We introduce primed-PCA (pPCA), an extension of the recently proposed EigenGame algorithm for computing principal components in a large-scale setup. Our algorithm first runs EigenGame to get an approximation of the principal com- ponents, and then applies an exact PCA in the subspace they span. Since this subspace is of small dimension in any practical use of EigenGame, this second step is extremely cheap computationally. Nonetheless, it improves accuracy significantly for a given computational budget across datasets. In this setup, the purpose of EigenGame is to narrow down the search space, and prepare the data for the second step, an exact calculation.

Speaker: Bálint Máté (UNIGE)

16:14 Local and correlated studies of humidity-mediated ferroelectric thin film surface charge dynamics

Electrochemical phenomena in ferroelectrics are of particular interest for catalysis and sensing applications, with recent studies highlighting the combined role of the ferroelectric polarisation, applied surface voltage and overall switching history.
Here, we present a systematic Kelvin probe microscopy study of the effect of relative humidity and polarisation switching history on the surface charge dissipation in ferroelectric Pb(Zr$_{0.2}$Ti$_{0.8}$)O$_3$ thin films. We analyze the interaction of surface charges with ferroelectric domains through the framework of physically constrained unsupervised machine learning matrix factorization, Dictionary Learning, and reveal a complex interplay of voltage-mediated physical processes underlying the observed signal decays. Additional insight into the observed behaviours is given by a Fitzhugh-Nagumo reaction-diffusion model, highlighting the lateral spread and charge passivation process contributors within the Dictionary Learning analysis.

Speaker: Iaroslav Gaponenko (DQMP, University of Geneva)

16:15 Machine learning applied to Gamma Ray Astronomy

Imaging atmospheric cherenkov telescopes are equipped with ultra high speed and single photon sensitive camera. Any extensive air shower will leave a signal of short duration (~10-20 ns) in the camera which contains the information about the primary particle type, energy and direction.
Since many years, this signal is transformed into an image from which a spatial parametrisation (Hillas parameters) is performed. These parameters are fed to boosted decision trees and more recently random forests to be able to discriminate gammas from protons, reconstruct their energies and eventually their directions.
The advent of both convolutional neural networks for image recognition and powerful computing ressources open a new era in gamma ray analysis. The particle type, energy and direction are now inferred from calibrated images without prior spatial parametrisation and provide, on simulation data, much better results than traditional analysis.
Based on this observation, the multi-messenger astronomy group at DPNC has developed several analysis based either on auto-encoders for data volume reduction, 1D CNN for photon stream extraction and more recently anomaly detection an CNN for trigger purposes. These works will be presented in the proposed poster.

Speaker: Matthieu Heller (DPNC - Université de Genève)

Poster_AI_in_Phys_Final.pdf

16:16 Faster Calorimeter Simulation

Faster Calorimeter Simulation

Speaker: Atul Kumar Sinha (UNIGE)

16:17 Draw the CURTAINs, Funnel, DREAM

A novel data driven anomaly detection method for extending bump hunts in high energy physics (CURTAINs), a novel likelihood based method that can be used to efficiently extend flows to high dimensional data such as that required in high energy physics (Funnels), and a tool for automating the validation of physics simulators (DREAM).

Speaker: Samuel Klein

CurtainsFunnelDREAM.pdf

16:18 Autoencoding star formation from cosmological parameters

As part of an interdisciplinary research collaboration, the summer project aimed at addressing the question of life in the multiverse by using a variational autoencoder to relate cosmic star formation to the underlying cosmological parameters.  The computational method is presented here. In future research, this will allow for the compression of $n$ cosmological parameters to a low-dimensional representation and the notion of equivalence of universes with seemingly different conditions.

Speaker: Mr Benjamin Hertzsch (ETH Zurich)

Geneva_AI_presentation.pdf

16:19 Estimating physical properties of galaxies using deep learning

Understanding the star formation and morphological properties of galaxies as a function of cosmic epoch is a critical exercise in studies of galaxy evolution. Traditional approaches to characterize these properties require a significant amount of human intervention and are computationally expensive, particularly for large samples of galaxies. Deep learning models have the potential to overcome these shortcomings. In this work, we present two deep learning models to predict : 1) stellar mass, star formation rate, and dust luminosity 2) r-band bulge-to-total luminosity ratio (B/T) of nearby galaxies. Our models are able to give an accurate estimation of these properties significantly faster than traditional approaches. The proposed deep learning approach could potentially save a tremendous amount of time, effort, and computational resources, particularly in the era of next-generation sky surveys such as the Legacy Survey of Space and Time (LSST) and the Euclid sky survey which will produce extremely large samples of galaxies.

Speaker: Omkar Bait (Observatory of Geneva)

PhyAI-day.pdf

16:20 ML Roulette for Track Overlay in ATLAS

Pileup modelling is an integral part in the data simulation and the downstream experimental analyses. Traditionally this is done by digitising the simulated pile up (PU) hits which are overlaid on digitised Hard Scatter (HS) hits, before reconstruction. A faster alternative is to reconstruct the PU separately and overlay them on HS tracks - i.e. overlay is done at the reconstruction level. This work presents a technique to identify regions where one can do the overlay without significant performance losses downstream, and provides a template for combining the two approaches for optimal speed and efficiency.

Speaker: Debajyoti Sengupta (Universite de Geneve (CH))

16:21 Graph Network Autoencoder for Jets

A significant number of tasks at ATLAS involve the generation or encoding of particle flow jets. These are jets which are represented by their constituent particles as a set or attributed point cloud. Such tasks include anomaly detection and simulation. Computation on these point clouds using standard dense networks is often difficult due to the requirement of some sort of ordering. We demonstrate that a graph network autoencoder provides two major benefits. Firstly, all operations within the network and it's training are invariant to any permutation on the set, including the calculation of the loss function. Secondly, the attention weighted message passing operations within the network allow us to focus on local information and relationships. We also present a novel batched autoregressive method for graph generation

Speaker: Matthew Leigh (University of Geneva)

16:22 Information-theoretic stochastic contrastive conditional GAN for physical data generation

Conditional generation is a subclass of generative problems when the output of generation is conditioned by a class attributes’ information. In this paper, we present a new stochastic contrastive conditional generative adversarial network (InfoSCC-GAN) with explorable latent space. The InfoSCC-GAN architecture is based on an unsupervised contrastive encoder built on the InfoNCE paradigm, attributes' classifier, and stochastic EigenGAN generator.
We propose two approaches for selecting the class attributes: external attributes from the dataset annotations and internal attributes from the clustered latent space of the encoder. We propose a novel training method based on a generator regularization using external or internal attributes every $n$-th iteration using the pre-trained contrastive encoder and pre-trained attributes’ classifier. The proposed InfoSCC-GAN is derived from an information-theoretic formulation of mutual information maximization between the input data and latent space representation for the encoder and the latent space and generated data for the decoder. Thus, we demonstrate a link between the training objective functions and the above information-theoretic formulation. The experimental results show that InfoSCC-GAN outperforms vanilla EigenGAN in image generation on several popular datasets, yet providing an interpretable latent space. In addition, we investigate the impact of regularization techniques and each part of the system by performing an ablation study. Finally, we demonstrate that thanks to the stochastic EigenGAN generator, the proposed framework enjoys a truly stochastic generation in contrast to vanilla deterministic GANs yet with the independent training of an encoder, a classifier, and a generator.

Speaker: Vitaliy Kinakh

InfoSCC-GAN-poster-november-3.pdf

16:23 Turbo-Sim: a generalised generative model with a physical latent space

Contact: guillaume.quetant@unige.ch

We present Turbo-Sim, a generalised autoencoder framework derived from principles of information theory that can be used as a generative model. By maximising the mutual information between the input and the output of both the encoder and the decoder, we are able to rediscover the loss terms usually found in adversarial autoencoders as well as various more sophisticated related models. Our generalised framework makes these models mathematically interpretable and allows for a diversity of new ones by setting the weight of each term separately. The framework is also independent of the intrinsic architecture of the encoder and the decoder thus leaving a wide choice for the building blocks of the whole network. We apply Turbo-Sim to a collider physics generation problem: the transformation of the properties of several particles from a theory space, right after the collision, to an observation space, right after the detection in an experiment.

Speaker: Guillaume Quétant (University of Geneva (CH))

TurboSim_FastCaloSim.pdf

16:24 Preprocessing solar spectra with a variational autoencoder to obtain the optimal dataset for solar flare prediction

Solar flares are the most energetic events in our solar system. Interacting with Earth’s magnetic field they may cause beautiful aurora or damage our modern infrastructure, but so far they cannot be reliably predicted. Novel methods such as neural networks allow us to cluster data, find rare events, or train models to attempt the prediction of solar flares. It is crucial to pre-process data such that the model can learn the relevant aspects of a phenomenon. A problem of flare data sets is that they always also contain non-flaring pixels, which may negatively affect any prediction models. We present a method using a variational autoencoder VAE to pre-process solar spectra and clean observations from Quiet Sun spectra which have no predictive power for solar flares.

Speaker: Jonas Zbinden (Université de Genève)

AIPhysics_day_V0.pdf

16:25 Gaia the 2 billion-star mission

A general overview poster of what the Gaia ESA spacecraft will deliver and the role of Switzerland in the variability processing & analyses. It mainly shows the increasingly larger number of sources we will classify in the different Gaia public data releases.

Speaker: Berry Holl (University of Geneva, Geneva Observatory)

GaiaThe2BillionStarEsaMission.pdf