The idea behind this meeting is to bring together those people interested in and working on machine learning, deep learning and artificial intelligence methods on DESY Campus, to get an overview how the heterogeneous mixture of scientists makes use of the same tools. As in the previous years this meeting shall strengthen the network of AI-experts working door to door here on campus in sometimes entirely different domains using the same tools and algorithms. This meeting explicitly addresses colleagues from all institutions on campus, from DESY to EMBL, from CSSB to XFEL and from MPSD to Hereon and beyond. Please feel invited!
The meeting will most likely be entirely virtual but eventually even in hybrid form. Please register to get all most recent information regarding the place and the video coordinates.
The deep learning activities of the Helmholtz Imaging Team at DESY IT comprise projects including semantic and instance segmentation, as well as object detection and localisation. Beyond the actual research on the DL methods, we aim to make such methods available online as scientific cloud compute services.
We are exploring different application of artificial intelligence, mostly aimed at simplifying and improving quality of data analysis, as well as at assisting with preventive identification of failures and with increased automation. Here we introduce some ideas.
Within the project "MDLMA" we work on generic techniques to automate the post-processing of images acquired on DESY beamlines and other facilities. One key challenge is the question of how to cope with the high memory demand on large 3D image data when training end-2-end architectures for transfer purposes?
AI has great potential to help with analysis of data generated at facilities such as Petra-III, Flash and EuXFEL. However, the range of experiments and speed of change of experiments in photon science facilities also pose major challenges for implementing AI - experimental techniques are diverse and there is rarely an experiment that does not change weekly. We will describe the areas where AI can potentially make a contribution and efforts in the area so far.
The fit between three-dimensional atomic models of biological macromolecules (think: DNA or corona spike protein) and experimentally measured data from synchrotrons, XFEL, neutrons and electron microscopy they are built on is estimated at only 80%. The Thorn lab uses neural networks to improve this fit through better ways to build these atomic structures and by finding errors in the experimental data.
We use third-party AI-based methods such as AlphaFold for modeling structures of protein complexes. We are also starting developing our own machine learning methods for applications in cryo-electron tomography, structural modeling, and systems biology. New to the AI field, we are interested in interactions with AI communities and collaborations.
This talk will give an overview of the current and planned GPU hardware on Maxwell as well as an outlook on what is to expect from AMD's, Intel's and NVIDIA's next generation of accelerator cards for AI/ML.
This will include benchmarks, timelines and an opportunity to ask questions.
Canceled
In the CMS experiment of the LHC, the Level-1 Trigger performs real-time event filtering of the particle collisions using FPGAs at a 40 MHz rate with a latency of several micro-seconds. For the HL-LHC, the L1T will be upgraded to cope with the increased luminosity and higher granularity of the data, enabling the use of Machine Learning in the FPGA algorithms.
We will briefly present the ongoing studies of the UHH CMS group related to ML-based classification of event topologies and jet identification in the Level-1 Trigger for the HL-LHC. First results and demonstrations will be presented, showing the proof-of-concept of such a novel approach in hardware trigger systems.
Longitudinal Phase Spaces (LPS) of electron bunches provide critical information for tuning and optimizing of FEL facilities. Here, we leverage the power of artificial intelligence to build neural network models using a data-driven approach, in order to bring destructive LPS diagnostics online and create a digital twin for the real machine. We will briefly discuss the motivation as well as the technical challenges. Furthermore, we will present the latest experimental results at the European XFEL and FLASH.
The recurrent task of manipulating the transversal beam parameters in the ARES experimental area poses as a test bed for studying reinforcement learning applications helping to automate complex tasks during accelerator operation. In this talk the current status and preliminary results of these studies will be presented.
Understanding dynamical features in time-resolved x-ray absorption spectra (TRXAS) can be quite challenging, even with theoretical simulations, due to high dimensionality of the data. In this talk, I will show how we can use machine learning techniques, in particular partial least squares regression, to reduce the dimensionality and get a single collective coordinate from ab initio trajectory data that most influences its TRXAS.
Automatic particle picking from 3D Cryo-EM tomograms still remains a challenging problem. We have developed a Deep learning based particle picker, CryoLearn. CryoLearn employs multiple layers of 3D Convolutional Neural Networks (CNN's). The CNN model is being trained on 3D Cryo-EM tomograms along with the coordinates of the particles residing in the tomogram annotated by human experts. The trained CNN along with a clustering algorithm is capable of predicting the coordinates of particles in an un-annotated 3D Cryo-EM map.
Detector simulation is a key cornerstone of modern high energy physics. Traditional simulation tools consume significant computational resources and are projected to be a major bottleneck at the high luminosity stage of the LHC and for future colliders. Deep generative models hold promise as a potential solution, offering drastic reductions in compute times. We present progress towards accurate simulation of particle showers in highly granular calorimeters in two directions. Firstly, initial progress on accurately simulating hadronic showers using a Wasserstein-GAN (WGAN) and a Bounded Information Bottleneck Autoencoder (BIB-AE) is demonstrated.
This talk will present a short overview of some of the Machine Learning (ML) activities from the CMS group at DESY.
The key features of the ML techniques will be highlighted, along with a few preliminary results.
We introduce a collection of datasets from fundamental physics research for supervised machine learning studies. These datasets are made public to simplify future work on cross-disciplinary machine learning and transfer learning in fundamental physics. Based on these data, we present a simple yet flexible graph-based neural network architecture that can easily be applied to a wide range of supervised learning tasks. An interface for the shared datasets and proposed algorithms can be found on GitHub: https://github.com/erum-data-idt/pd4ml
https://cloud.tuhh.de/index.php/s/QMk8mpam7wsPa5x