We are looking forward to welcoming you to the first HAMLET-PHYSICS Conference/Workshop, to be held in sunny Copenhagen, August 19 - 21, 2024.
The workshop has three main goals: 1. To bring together Danish and international physicists using ML to meet, share ideas, and build community across location and physics specialty 2. To bring domain scientists into close contact with ML experts, to build community across the theory - application bridge 3. To provide a friendly environment for researchers to share best practices, for students to interact with experts, and for other sciences and industry to understand the state of ML in physics |
![]() |
Confirmed speakers include Thea Aarrestad (ETH Zurich), Savannah Thais (Columbia University) and Roman Pasechnik (Lund University).
Abstracts are open for contributions at the intersection of machine learning and
This is not an exhaustive list. We warmly welcome all submissions for talks, suggestions and ideas, and will strive to accommodate all submissions.
Abstracts submitted after the deadline will be considered on a case-by-case basis.
Monday August 19th will feature a poster session and reception event at the University of Copenhagen Biocenter.
On Tuesday evening August 20th, the workshop will take to the rails: A heritage 1950s Norwegian State Rail diesel locomotive will take workshop attendees from Copenhagen (Østerport Station) to Kronborg Castle in Helsingør (location of Shakespeare's Hamlet tale).
While on board, refreshments will be served, and breakout sessions will occur according to attendees research areas of interest. A visit of Kronborg will be included, along with an open-air conference dinner in Helsingør.
The Standard Model of particle physics is an effective theoretical model remarkably describing matter particles and forces at the fundamental level. Despite its great success, it leaves many open questions such as the nature of Dark Matter and the origin for matter-antimatter asymmetry in the Universe.
The search for New Physics is a great challenge of contemporary science, which nonetheless has not been hitherto successful in recovering detectable signatures that could be distinguished from noise backgrounds in multi-dimensional datasets. Within this context, Machine Learning techniques have been deployed to overcome phenomenological challenges in the hunt for new particles.
The discovery of the Higgs boson has become perhaps one of the most striking example of the successful use of methods borrowed from Artificial Intelligence in high energy particle physics.
In this talk, I will overview the state-of-the-art, impeding issues and recent progress in development of advance Machine Learning methods and their applications in the ongoing searches for subtle new phenomena in fundamental physics.
GraphNeT, a deep learning framework for neutrino telescopes, provides a common library for deep learning experts and neutrino physicists in need of advanced deep learning techniques for their analyses. GraphNeT is a detector agnostic framework, making it easy and intuitive to apply deep learning techniques from one experiment to another, and to export models to the physicists that relies on them. GraphNeT lowers technical barriers, enabling physicists to utilize deep learning without extensive expertise, and sets the stage for inter-experimental collaboration. This poster presentation highlights the key developments of GraphNeT 2.0 and their impact on how they provide a way for neutrino telescopes to work together on advancing the state-of-the-art for deep learning based techniques in the field.
We pursue the use of deep learning methods to improve state-of-the-art computations in theoretical high-energy physics. Planar N = 4 Super Yang-Mills theory is a close cousin to the theory that describes Higgs boson production at the Large Hadron Collider; its scattering amplitudes are large mathematical expressions containing integer coefficients. In this paper, we apply Transformers to predict these coefficients. The problem can be formulated in a language-like representation amenable to standard cross-entropy training objectives. We design two related experiments and show that the model achieves high accuracy (> 98%) on both tasks. Our work shows that Transformers can be applied successfully to problems in theoretical physics that require exact solutions. (Based on arXiv:2405.06107 [cs.LG])
Artificial intelligence (AI) and machine learning (ML) have become critical tools in many scientific research domains. In many areas of physics ML tools are essential to meeting the computing needs of current and future experiments and to ensuring robust data reconstruction and interpretation. In addition to being powerful tools for scientific research, ML and AI are now ubiquitous in nearly all facets of society, from healthcare to criminal justice, from education to transportation. These applications have the potential to address critical community needs and improve educational, health, financial, and safety outcomes; however, they also have the potential to exacerbate existing inequalities and raise concerns about privacy, surveillance, and data ownership.
In this talk I will explore critical ethical considerations that arise when designing, developing, and deploying data science and ML/AI systems when modeling both scientific and social systems, including bias and fairness, task design, equitable data practices, model evaluation, and more. These topics are essential to ensuring we can trust the results of our AI models in our scientific research and to helping us build a more complete understanding of how AI models behave and when we can appropriately use them to make or support decisions. Although these are not purely mathematical problems that can be fully solved with technological approaches, I will emphasize quantitative approaches and techniques to address these issues and discuss what unique roles physicists might play in co-creating a more just and responsible future of technology. I will show examples of recent work from particle physics and astronomy that exhibit how physics data can be used to develop AI models that both enhance the scientific capacity of our experiments and help us gain foundational insight into AI.
Detection of neutrinos at ultra-high energies (UHE, E >$10^{17}$eV) would open a new window to the most violent phenomena in our universe. However, owing to the expected small flux of UHE neutrinos, the detection rate will be small, with just a handful of events per year, even for large future facilities like the IceCube-Gen2 neutrino observatory at the South Pole. In this contribution, I will discuss how to substantially enhance the science capabilities of UHE neutrino detectors by increasing the detection rate of neutrinos and improving the quality of each detected event, using recent advances in deep learning and differential programming. First, I will present neural networks replacing the threshold-based trigger foreseen for future detectors that increase the detection rate of UHE neutrinos by up to a factor of two. Second, I will outline and present preliminary results towards an end-to-end optimization of the detector layout using differential programming and deep learning, which will improve the neutrino direction and energy determination. I will present new results of a deep, ResNet-based neural network that, combined with Normalizing Flows, predicts the full Posterior PDF of neutrino direction and energy. The expected improvements are estimated to be equivalent to building an up to three times larger detector, accelerating the research field by a decade.
In a recent paper we presented a new benchmark tool for symbolic regression and tested it with cosmology datasets. This new benchmark tool is easy to use, and we would like to spread the word and encourage other people to try it in their respective field where applicable. Out of the box it has ten different machine learning algorithms, but we also encourage the community to add their own methods to the framework and expand the capabilities. In this talk I will discuss how it works and the paper we published with a test application in cosmology. We find no indication that performance of algorithms on standardized datasets are good indications of performance on cosmological datasets. This suggests that it is not necessarily prudent to choose which symbolic regressions algorithm to use based on their performance on standardized data. Instead, a more prudent approach is to consider a variety of algorithms. Overall, we find that most of the benched algorithms do rather poorly in the benchmark and suggest possible ways to proceed with developing algorithms that will be better at identifying ground truth expressions for cosmological datasets. As part of this publication, we introduce our benchmark algorithm cp3-bench which we make publicly available at https://github.com/CP3-Origins/cp3-bench.
Black holes hold a tremendous discovery potential, and experiments such as the Event Horizon Telescope and its next generation upgrade could provide important cues about their structure. I will present a pedagogical discussion on the most relevant features in black-hole images (e.g. photon rings) and their detectability using realistic telescope arrays. I will focus on the structure of Fourier-plane data when expressed in terms of standard interferometric variables known as closure quantities, and motivate the use of machine learning to identify specific features and quantify observational uncertainties.
In this presentation, we will explore the application of advanced machine learning techniques in cosmology, focusing on the analysis of Cosmic Microwave Background (CMB) maps. Accurately calculating the tensor-to-scalar ratio from CMB data is a crucial yet challenging task, as it holds the key to understanding primordial gravitational waves and the early universe's inflationary period. I will discuss the use of deep learning and other sophisticated algorithms to extract meaningful features and parameters from CMB maps, which can be readily reapplied in other areas of cosmology and astrophysics studies. These methods offer robust tools for dealing with the complexities and high-dimensional nature of data. By leveraging machine learning, we can enhance our ability to simulate, analyze, and interpret CMB observations, providing deeper insights into the universe's fundamental properties. The versatility and potential of machine learning in advancing our understanding of the cosmos will be highlighted, by showing data analysis techniques applicable in all scientific disciplines.
The search for Earth-like planets around Sun-like stars using Doppler radial velocity (RV) measurements is challenged by stellar variability. Stellar variability alters the shapes of spectral lines, introducing spurious RVs that obscure the true RV signals of planets.
To address this issue, we study the Sun as a star, where we know the ground truth of the solar RVs as a result of solar variability. Our experiment begins by extracting line shape information from the "Sun-as-a-star" observations using FIESTA (Zhao et al., 2022), a Fourier-based parameterization of the spectral cross-correlation function. We then feed these extracted features into a 2D convolutional neural network to predict the resulting solar RVs. Furthermore, this model can be transferred to other Sun-like stars for predicting their stellar variability RVs in general.
This method will help us separate the stellar RV noise from the planetary RV signals, which is essential for accurately detecting exoplanets amidst stellar noise.
Recently proposed Neural-LAM forecasting models (https://arxiv.org/abs/2309.17370, https://arxiv.org/abs/2406.04759) bring state-of-the-art graph-based approaches to Limited Area Modeling (LAM). The models generate and use an area-restricted graph and take forcing inputs from a host model to handle the boundary conditions. Initial promising results have motivated interest in developing the approach further. To this end, a cross-national collaborative community has been formed (https://github.com/mllam). We report ongoing work from this community to 1) survey available and relevant reanalysis datasets for training, 2) transform these source datasets to common zarr-based training-ready datasets, on 3) preliminary results applying the Neural-LAM to new spatial domains, and on 4) research questions around global-modal coupling and ensemble forecast predictions currently planned to be tackled within the community. Within the community we are developing a common tooling and infrastructure to share training datasets, and working on shared verification tooling targeting metrics relevant for kilometre-scale forecasts, both of these developments we will detail. The flexible and modular codebase for training and evaluating LAM models is available in an open source repository that welcomes contributions.
Information processing and analysis of time series are crucial in many applications but often face constraints such as high computational complexity. Quantum reservoir computing, which combines a reservoir of neuromorphic quantum hardware with a simple neural network, offers a promising solution. By utilizing the high-dimensional space and dynamics of quantum systems, this approach enables the creation of models that can handle more challenging temporal learning tasks.
In this project, we simulate a system of interacting coupled quantum dots connected to electronic leads as a quantum reservoir. We use a Lindblad master equation approach to calculate the dynamics and transport through the systems and evaluate its performance through several benchmark tests.
Geomagnetic storms, resulting from solar activity, pose significant risks to satellite operations, communication systems, and power grids. Accurate forecasting of these storms is crucial for mitigating their impacts. As the final project in the course "Applied Machine Learning" at the University of Copenhagen, we explore the application of convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to forecast geomagnetic storms using satellite image data from the Solar Dynamics Observatory (SDO). By leveraging solar images capturing phenomena such as solar flares and coronal mass ejections (CMEs) in the 171Å band, our neural network models are trained to identify patterns and temporal sequences indicative of the geomagnetic activity. Preliminary results demonstrate that the neural networks work well for geomagnetic forecasting on short timescales. Future work should focus on extending the models for predictions further into the future and perhaps also more specifically optimizing the models for geomagnetic storm prediction, if this is desired.
The James Webb Space Telescope is helping astronomers to push back the frontiers of observability. This is particularly true for exoplanets, as we are now looking at small, terrestrial planets, searching for hints of an atmosphere. These signals are very small, and as with every new instrument, most systematics (and their source) remain very poorly understood. In an effort to investigate possible patterns in the JWST/MIRI detector responsible for the correlated noise observed in the light-curves, I apply clustering techniques on each pixel's time series information. The results suggest that 1) There are no major gradients or positional patterns across the detector, 2) Cosmic rays/bright polluters "shock" individual pixels beyond their occurrence integration and 3) There seems to be at least one high frequency signal (whose source remains unknown), possibly polluting the light curves.
Accurate representations of unknown and sub-grid physical processes through parameterizations (or closure) in numerical simulations with quantified uncertainty are critical for resolving the coarse-grained partial differential equations that govern many problems ranging from weather and climate prediction to turbulence simulations. Recent advances have seen machine learning (ML) increasingly applied to model these subgrid processes, resulting in the development of hybrid physics-ML models through the integration with numerical solvers. In this work, we introduce a novel framework for the joint estimation of physical parameters and machine learning parameterizations with uncertainty quantification. Our framework incorporates online training and efficient Bayesian inference within a high-dimensional parameter space, facilitated by differentiable programming. This proof of concept underscores the substantial potential of differentiable programming in synergistically combining machine learning with differential equations, thereby enhancing the capabilities of hybrid physics-ML modeling.
FoCal-H is a hadronic calorimeter currently under development and is intended as a part of an upgrade to the ALICE experiment at the LHC at CERN. As part of an Applied Machine Learning course, we looked at real data from FoCal-H and used machine learning techniques on it. Data was collected during testbeams where particle type and energy are selected and fired at the device. With labelled data we were able to do supervised learning as well as unsupervised learning. Our approach was to compute features for all the events, and in particular, we computed Hu Moments, which are moments that are invariant under certain transformations. We were able to train models to predict particle type with around 98% accuracy, and less successful with other labels like energy. We also did anomaly detection by hand-picking good events and translating, rotating and mirroring them into new events to augment the original hand-picked set. This method showed promise.
The advent of gravitational wave astronomy has enabled the direct study of the synthesis of heavy elements such as rare earths, lanthanides, and actinides. This was exemplified by the kilonova AT2017gfo/GW170817, where several heavy element emission lines were confirmed. The primary challenge in advancing this research is the lack of accurate and complete atomic data for these elements in ionization stages I-IV.
Our project, part of the ERC-funded HeavyMetal collaboration, aims to address this by using the GRASP atomic structure code to calculate theoretical atomic data. GRASP, known for its accuracy, requires extensive manual input and is time-consuming. The process involves iteratively adding Configuration State Functions (CSFs) and running GRASP to check for convergence. Each additional CSF can improve accuracy but also significantly increases computational time, making it critical to determine the optimal set of CSFs.
We propose to use machine learning to automate this iterative decision-making process. We will develop a transformer-based surrogate model to predict the convergence of GRASP for a given set of CSFs and the improvement in output accuracy. This model will enable a reinforcement learning esque approach to efficiently explore CSF combinations, optimizing for both accuracy and computational time.
This project will significantly reduce the manual effort required for GRASP calculations and use computational resources more efficiently, through the use of machine learning, enabling more rapid and comprehensive studies of the heaviest elements synthesized in neutron star mergers.
Our understanding of past climate and the mechanisms that drive climate change can be improved through analysis of the excellent palaeoclimatic data available in the Greenland ice cores, but such insights depend on having an established chronology of the ice core. Therefore, the dating of ice cores is an essential part of palaeoclimatic science. However, this dating is often carried out by manual identification, which is both time-consuming and somewhat subjective.
A GRU Encoder-Decoder model for automatic annual layer identification is developed. The GRU provides a sigmoid output, which is then used to find annual layer positions with a peak detection algorithm. The method is applied to the ice cores NGRIP, GRIP and DYE-3 using the manually identified layer positions from GICC05 and GICC21 as targets in training. These annual layer positions, together with reference horizons from stratigraphic markers, are used for evaluating predictions from the model.
The developed GRU model is found to be able to match the GICC annual layer count for all three tested ice cores within a difference of $4.36\%$. Within the reference horizons used for GRIP and NGRIP, upwards of $78.9\%$ of GRU counts either agree or are within a margin of $\pm 1$ of the number of annual layers identified in GICC05. The proposed GRU model is useful for validating timescales, since dubious annual layers can be located by examining the difference between existing annual layer positions and the ones provided by the model.
This study aims to determine glacier thickness based on observational data. Due to global warming the rising temperatures leads to increased melt from land-based bodies of ice. Increased melt from glaciers both pose a risk to societies that are based along rivers produced from glacier discharge, as well as having direct impact on sea level rise. Determining the thickness or the mass of these glaciers help to better predict increased river activity and sea level rise.
In this study, three different Machine Learning algorithms were used. First a Boosted Decision Tree and Neural Network model were implemented based on field measurements, to compare the performance based on the mean absolute error. Secondly, a Convolutional Neural Network were implemented on the field measurements and satellite images to test if the additional observations would improve the model.
From this analysis we found that the Boosted Decision Tree, XGBoost, had the lowest mean absolute error and thus was the best at predicting glacier thickness. Using the Convolutional Neural Network with satellite images, did not improve the precision compared to the other models. Further work could include adding additional features to the data, such as latitude and longitude to the satellite images and surface temperature at the location of the glacier to increase the precision of the prediction models.
Star formation is a multi-scale problem. Theoretically, only global simulations that account for the connection from the large-scale gas flow to the accreting protostar reflect the observed complexity of protostellar systems. In such global models, star are born in a stochastic process as a statistical ensemble and it is not possible to create constrained models that match specific observations ab initio. Observationally, a combination of single dish and interferometers are able to resolve the nearest protostellar objects on all scales from the protostellar core to the inner few AU, for the nearest protostars. It is challenging to create models that objectively and non-parametrically match the observed properties, and henceforth can be used to better interpret the data and understand the underlying physical mechanisms.
We have developed a new methodology for using high-resolution models and post-processing methods to match simulations and observations non-parametrically using deep learning in a semi-automatic fashion and extract robust physical indicators which can radically speed up the interpretation of high-resolution interferometric high-resolution observations. With machine-learning, we can down-select from large datasets of synthetic images to a handful of matching candidates. This is particularly useful for binary and multiple stellar systems and has the end goal to infer the underlying physics that drives the creation and evolution of protostellar systems.
Suggestion locations:
- Sunny: Broens Gadekøkken (Metro Line 3, stop Kongens Nytorv, walk 300m)
- Rainy: Meatpacking District (Metro Line 3, stop København H, walk 200m)
At the CERN Large Hadron Collider (LHC), real-time event filtering systems must process millions of proton-proton collisions every second on field programmable gate arrays (FPGAs) and perform efficient reconstruction and decision making. Within a few microseconds, over 98% of the collision data must be discarded fast and accurately. As the LHC is upgraded to its high luminosity phase, HL-LHC, these systems must deal with an overwhelming data rate corresponding to 5% of the total internet traffic and will face unprecedented data complexity. In order to ensure data quality is maintained such that meaningful physics analyses can be performed, highly efficient ML algorithms are being utilised for data processing. This has necessitated the development of novel methods and tools for extremely high throughput, ultra low latency inference on specialised hardware.
In this talk, we will discuss how real-time ML is used to process and filter enormous amounts of data in order to improve physics acceptance. We will discuss state-of-the-art techniques for designing and deploying ultrafast ML algorithms on FPGA and ASIC hardware. Finally, we will explore applications of real-time inference in particle physics experiments and discuss how anomaly detection can be used to discover New Physics.
We present findings from the FASTSUM collaboration on the behaviour of bottomonium particles (bound states of a bottom quark and its antiquark) at different temperatures. To analyze this, we used three methods: a Maximum Likelihood approach with a Gaussian function for the ground state, the Backus-Gilbert method, and a machine learning technique called Kernel Ridge Regression. Our study uses lattice simulations with 2+1 quark flavours, covering temperatures from 47 to 375 MeV. The findings help us understand how bottomonium particles behave in hot environments, which is important for studying the properties of quark-gluon plasma and the dynamics of heavy quarks in extreme conditions.
Accurately representing the tropical sea surface temperature (SST) remains a significant challenge for general circulation models. One of the largest sources of uncertainty is the vertical turbulent mixing. To accurately represent the distribution of ocean mixed layer depths (MLD), turbulence closure schemes necessitate careful tuning. This is commonly done manually by comparing with mixed layer depth climatologies. Advancements in machine learning research introduce a new strategy: automated tuning. VerOpt, an add-on to the Python-based ocean model Veros, uses Gaussian processes to emulate an objective function in a multi-dimensional parameter space. We demonstrate how VerOpt can be used to search the joint parameter space of the vertical & horizontal mixing and air-sea flux parameterisations. Furthermore, we discuss the technicalities and advantages associated with using a python-based ocean model that utilises JAX for GPU acceleration.
The development of innovative methods for fission trigger construction is part of the FRØZEN project which aims to get a better understanding of the angular momentum generation and the energy partition between fragments in the fission process. The reconstruction of the very first moments after the scission point is essential and requires correlated neutron and gamma detection as well as measuring the kinematic properties of fission fragments. Such a measurement could be achieved thanks to the last generation of hybrid $\gamma$-spectrometer named $\nu$-Ball2, coupled to a double Frisch-Grid Ionisation Chamber (dFGIC). In this experiment, a spontaneous ${}^{252}$Cf fission source was used. However, for other fissioning systems that require the use of a primary beam, fission could become a minor nuclear reaction compared to other processes. Additionally, with the increasing size of nuclear physics experimental setup, the need to recognize rarer reaction mechanism, one of the main challenges nowadays in nuclear physics is to develop more and more selective data analysis methods for more and more contaminated datasets.
AI models are being developed to replace the usual data analysis techniques for reconstructing the fission events, exploring the limits of AI implementation in such context. The first implementation is deeply motivated by the computationally expensive and time-consuming characteristics of more usual trace (sampled signal) analysis approaches, currently used to analyze the dFGIC response and tag fission for $\nu$-Ball2 setup. Some promising regression and convolutional neural network models have been tested to obtain precise fission tag time, the deposited energy, and the electron drift time from the dFGIC traces. The second implementation tackles the challenge of recognizing fission events from a polluted dataset by developing an AI-based algorithm to recognize fission solely based on the $\nu$-Ball2 response function. The fission fragments de-excitation process is reconstructed from the correlations between the individual fission fragments pairs and observables, such as gamma and neutron energies and multiplicities. AI models can be used to evaluate the impact of each observable into identifying fission. Once the algorithm is trained, it could be applied to various fissioning systems without the need for an ancillary fission tag detector, such as the dFGIC.
In the field of neutron scattering experimentation, the use of complicated sample environments containing strong magnets introduces significant challenges. The inclusion of substantial material within the experimental structure is necessary in order to withstand the large magnetic forces, which eventually influences the experimental outcome through events of multiple neutron scattering. This project leverages data science and machine learning techniques to investigate and predict background scattering due to multiple scattering that takes place in the complex sample environment.
Utilizing the McStas neutron ray-trace simulation package, a detailed model of the 15 T magnet for the BIFROST spectrometer at the European Spallation Source (ESS) was developed. The model was parameterised to cover different experiment setups with a number of simulation parameters, generating a substantial amount of simulation results. A comprehensive database of 24000 simulation results was constructed, and analysed to uncover underlying patterns and relationships between the experimental setup parameters and the observed background scattering. Subsequently, machine learning models were trained, fine-tuned and tested, and their predictive accuracy was assessed.
This project contributes to the field of neutron scattering by providing a novel approach to addressing the challenges of background prediction due to multiple scattering in complex sample environments and can serve as an introduction to a new method of background recognition, paving the way for the development of automated background prediction tools that can be used within a wide range of instruments, with combinations of simulated and experimental data in the future. It also exemplifies the potential of data science in advancing experimental physics.
Bayesian model selection is an invaluable tool when comparing different inflationary models. This is done by computing the Bayesian evidence for the different models using nested sampling, and then comparing them using the Bayes factor. Computing the Bayesian evidence of a model using an Einstein-Boltzmann solver code takes weeks, if not months for complex models, and it is therefore very rarely done. Using a neural network, e.g. from the CONNECT framework, to emulate cosmological observables, the time used to compute the Bayesian evidence of a model can be brought down from tens of thousands of CPU hours to about 100 CPU hours.
In this talk, I will present the setup and programs used to compute the Bayesian evidence using the neural network framework CONNECT. I will also show the Bayesian evidence for the cosmological model ΛCDM+ɑ_s+r computed with CLASS and CONNECT respectively. Furthermore, I will also present a comparison of different inflationary models using the Bayes factor.
While identification, characterisation, and triggering mechanisms of active galactic nuclei (AGN) have been since the 80's, the discussion has only been extended to include dwarf galaxies within the last decade.
This study aims to explore a novel AGN identification technique using a random forest (RF) classification technique, compare it to established identification methods, and investigate which set of properties/features constitute the best RF model.
Data is sourced from multiple catalogues: MaNGA (and its value added catalogue, Firefly) provides spatially resolved spectra of 10\,104 galaxies of which 1\,149 are dwarf galaxies. These galaxies constitute the base data set, and infrared (WISE) and X-ray (XMM) observations are matched to these. The NASA-Sloan Atlas is used for estimating environmental parameters.
The best model (from F1 score alone) is using internal features only of more massive galaxies. This model tends toward weighing fewer features higher and ignoring parameters that are less directly related to AGN ionisation. Conversely, this model disagrees the most with observations when it comes to dwarf galaxies, but provide twice as many dwarf AGN candidates as observations, and up to thrice as many compared to using intermediate mass galaxies as training set.
This approach provides a novel and interesting venue for identification of AGN in dwarf galaxies, but the method still requires fine tuning such as feature selection optimisation and validity assessment -- are the predicted AGN actually AGN? If so, RF can be used to increase the sample size of known dwarf AGN and to adjust observational diagnostic diagrams in the low mass regime.
Large Language Models (LLMs) have become extremely popular over the past year and a half. Researchers have experimented with using LLM-powered chatbots (usually Chat-GPT) as an “assistant” to help them understand existing work and even discuss new research ideas. However, a pure LLM chatbot tends to make plausible sounding but unsupported statements, and cite non-existing
papers, which severely limits its usefulness. Retrieval Augmented Generation (RAG) restricts an LLM chatbot to using only a particular body of literature to answer questions. The answer is extracted from retrieved text, and the sources (papers, sections, paragraphs) are listed. Therefore a RAG chatbot can greatly support literature search and research more generally.
The field of astrophysics is ideally suited for experimenting with this technique. The obstacles are limited, as scientific literature and technical documentation tend to be readily available, and the researchers have a high level of technical literacy and affinity. At the same time, the potential benefits are high, as researchers often need to familiarize themselves with new instruments and techniques.
In collaboration with the NLP team at ESA we are developing a RAG chatbot using Gaia manuals and science papers. Gaia data products are complex and will remain relevant for decades, making it important that helpful information remains easily accessible for astrophysicists. A working prototype can be accessed at https://gaiachat.streamlit.app. In parallel, we are working on a similar system for the upcoming PLATO mission (https://platochat.streamlit.app/).
We are exploring the potential of supporting students, as well as academia - industry collaborations, by developing a smart portal that helps students in finding well-fitting internship opportunities. Here there will be a central sign-up for industry partners, and students can then search for opportunities that match their interests and skills in dialog with the system. We expect that this will encourage more students to do an internship, leading to positive experiences for both students and industry partners, and increased collaboration.
Detailed knowledge of the radiation environment in space is an indispensable prerequisite of any space mission in low Earth orbit or beyond. The RadMap Telescope is a compact multi-purpose radiation detector that provides near real-time monitoring of the radiation aboard crewed and uncrewed spacecrafts. A first prototype has been deployed on the International Space Station in April 2023 for an in-orbit demonstration of the instrument’s capabilities. RadMap’s main sensor consists of a stack of scintillating-plastic fibres coupled to silicon photomultipliers. The perpendicular alignment of fibres in the stack allows the three-dimensional tracking of charged particles as well as the identification of cosmic ray ions by reconstruction of their energy-loss profiles. We trained a set of convolutional neural networks on simulated detector data to perform an event-by event reconstruction of track parameter, ion type and energy. In addition to our current offline analysis, we plan to implement the analysis framework on the instrument’s flight computer to analyze measurements without requiring the transmission of raw
data to Earth. In this contribution, we will describe our neural-network-based reconstruction methods and present first results.
Suggestion locations:
- Sunny: Broens Gadekøkken (Metro Line 3, stop Kongens Nytorv, walk 300m)
- Rainy: Meatpacking District (Metro Line 3, stop København H, walk 200m)
Suggestion locations:
- Sunny: Broens Gadekøkken (Metro Line 3, stop Kongens Nytorv, walk 300m)
- Rainy: Meatpacking District (Metro Line 3, stop København H, walk 200m)
Suggestion locations:
- Sunny: Broens Gadekøkken (Metro Line 3, stop Kongens Nytorv, walk 300m)
- Rainy: Meatpacking District (Metro Line 3, stop København H, walk 200m)
Symbolic regression (SR) is a machine learning technique that seeks to identify mathematical expressions that best fit a given dataset. This method is particularly beneficial in scientific research, due to its ability to generate interpretable models that provide insight into the underlying mechanisms, in sharp contrast with purely data-driven, black-box models. Interpretable machine learning is gaining importance as models are increasingly deployed in critical societal applications like finance and medicine. Also, models discovered by SR are prone to analyses that allow to assess their out-of-distribution behavior. These beneficial properties have led to the application of SR across various fields, including physics, biology, climate modeling, finance, and various engineering disciplines. Despite these opportunities, the application of symbolic regression faces several challenges. One major issue is the computational complexity associated to the combinatorial search within the space of mathematical expressions, which requires the use of high-performance computing, scalable algorithms, and smart strategies or integration of domain-specific knowledge to guide the search and make it more efficient. Additionally, ensuring robustness and generalization of the discovered models can be difficult, as overfitting to noisy data or limited datasets may occur. In this talk, we wish to stimulate the debate among researchers about both opportunities and challenges concerning the use of SR for Science.
Computer simulations aim to replicate physical phenomena as precisely as possible, but they often require extensive knowledge and rely on differential equations, which typically lack analytical solutions. The recent rise of Machine Learning, driven by large amounts of data, has offered an alternative to traditional human-designed methods. However, these data-driven approaches produce complex, black-box models that are hard to manage and interpret. To address this, Symbolic Regression (SR) has gained popularity for its ability to derive understandable mathematical equations from data. Nevertheless, most prior research only focused on re-discovering conventional physics equations or dynamical systems described by ordinary differential equations.
To further push the boundaries of the existing methods, we developed a framework that integrates SR with Discrete Exterior Calculus (DEC) for automated discovery of field theories. DEC provides a discrete geometric representation of physics, avoiding the need for differential formulations and employing a concise set of operators. This approach significantly reduces the SR search space, improving generalization even with limited data. We validated our framework by successfully re-discovering multiple Continuum Physics models using synthetic experimental data.
In High Energy Physics (HEP) experiments, such as those conducted at CERN, the demand for increasingly powerful data acquisition systems (DAQ) is a common necessity. These systems typically encompass a network of computers, high-end FPGAs, and in certain instances, even GPUs.
Despite their substantial capabilities, these multi-million-euro facilities often lie dormant when experiments are not in use, like during technical stops.
Our proposal introduces a framework designed to reclaim these idle clusters for executing Machine Learning algorithms such as Monte-Carlo generation or further data analysis.
This innovative approach aims to maximize the utilization of resources within these facilities. We will share our results at running an example model used for parameterizing the LHCb tracking resolution in a DAQ FPGA, along with the outcomes we achieved. We will also outline our plans for moving forward towards heterogeneous computing.
The Large-Sized Telescopes (LSTs) are one of three telescope types being built as part of the Cherenkov Telescope Array Observatory (CTAO) to cover the lower energy range between 20 GeV and 200 GeV. The Large-Sized Telescope prototype (LST-1), installed at the Observatory Roque de Los Muchachos, La Palma, is currently being commissioned and has successfully taken data since November 2019. The construction of three more LSTs at La Palma is underway. A next generation camera that can be used in these three upcoming LSTs is being designed. One of the main challenges for the advanced camera is the 1GHz sampling rate generating 72 Tbps of data. A filter removing events caused by random coincidences from the night sky background or sensor noise shall reduce the data rate to 24 Gbps, corresponding to an event rate of approximately 30 kHz. Finally, a software stereo trigger, featuring deep learning inference on a high-speed Field Programmable Gate Array (FPGA), will further reduces the event rate to < 10 kHz.
To achieve such a large reduction, the different trigger levels are currently being developed for the implementation on FPGAs. This presentation focuses on porting the final trigger stage, a real-time deep learning algorithm, to FPGAs using two different approaches: the Intel AI Suite and the hls4ml packages. We compare the trade-offs obtained in FPGAs against running it in GPUs.
A long baseline neutrino experiment, the purpose of the ESS Neutrino Superbeam+ (ESS$\nu$SB+) is to precisely measure $\delta_{CP}$ in the leptonic sector using a powerful proton beam produced at the European Spallation Source (ESS).
In this endeavor, accurate and fast event reconstruction is central for the design and performance of the ESSnuSB detectors. While precise, the currently proposed likelihood-based method is computationally expensive.
In this work, we investigate the use of Graph Neural Networks (GNNs) for classification of muon- and electron neutrino events in the Near Detector of the ESSnuSB experiment, as well as for classification of charged- and neutral current events and of the presence of pion production.
We demonstrate that the accuracy of the GNN method is comparable to that of the likelihood method, and that the GNN can even learn the signatures of, and accurately identify, complex events that are currently discarded, while providing a factor $10^4$ increase in reconstruction speed.
The standard Echo State Network (ESN) model has gained wide recognition for its success in modeling nonlinear auto-regressive (NAR) dynamical systems. For example, it is suitable for modeling the Mackey-Glass system, which is used for analyzing bifurcations in physiological applications. It is also employed to analyze the Lorenz system, which models several physical phenomena such as the dynamics of electro-rotation, thermohaline ocean flow circulation, the Malkus waterwheel, and weather transition dynamics. Furthermore, the ESN model is particularly interesting in a quantum computing context. The recurrent topology of the ESN model ensures that the non-linear transformation of the input history is stored in internal states. The system operates in a regime that can be either stable or can exhibit chaotic behavior. Over the past few years, several studies concerning the network stability have been carried out using Lyapunov exponents and the computation of the edge of chaos. These works have shown the significant impact of the eigenvalues on the system stability.
Recently, a framework was introduced where a graph embedding technique and evolutionary algorithms were combined to fully train a recurrent network with the ESN characteristics. This neuroevolutionary approach for optimizing an ESN works well in practical applications, but it may be expensive due to the numerous computations of the spectral radius of a weight matrix. In recurrent topologies computing the spectrum may be unstable and computationally expensive. In this work, we use an evolutionary framework to analyze the impact of the pseudo-spectrum on the stability of the ESN dynamics. The computation of the pseudo-spectrum is robust and cheaper than the computation of the eigenvalues. We carry out experiments showing the relationship between the pseudospectra and stability of the recurrent network on both MGS and Lyapunov systems.
This study explores various neural network approaches for simulating beam dynamics, with a particular focus on non-linear space charge effects. We introduce a convolutional encoder-decoder architecture that incorporates skip connections to predict transversal but also coupled 3D electric self-fields. The model demonstrates robust performance, achieving a Mean Absolute Percentage Error (MAPE) of $0.5\%$ within just a few minutes of training. Our findings indicate that these advancements could provide a more efficient alternative to numerical non-linear space-charge methods in beam dynamics simulations, where the speed up is significant.
Semiconductor nanowires are successfully used as a biosensing substrate due to their ability to enhance the fluorescence of bound fluorescently labeled molecules. This enhancement, which combines excitation, quantum yield, and collection enhancement, is influenced by nanowire diameter, refractive index, and the wavelength of the bound fluorophore. Combined with a large surface-to-volume ratio and high throughput in the field of view, nanowires allow the quantification of molecular concentrations as low as 100 fM and the observation of single-molecular binding processes over time [1, 2]. However, information about a bound molecule’s position along the nanowires’ z-axis has not been available to date.
Here we extend the capabilities of nanowire-based single-molecule detection to include information about the molecule’s position along the z-axis. For this purpose, we utilize the image formation asymmetries observed for fluorophore emission wavelengths where fluorescence enhancement is sigificant and depend on the fluorophore binding position (fig 1b). We used numerical solutions of Maxwell's equations to simulate fluorophore excitation and emission enhancement by nanowires, followed by image creation [3] (fig. 1 a-c). These axial position-dependent images were then used to train convolutional neural networks to predict the binding position with 50 nm step-size. The ability to simulate images for specific nanowires and microscopes enables these trained networks to be applied to real microscopy images. High prediction accuracies suggest that more advanced neural networks can be implemented to track single-molecules' motion along the nanowire axis, and even opening a possibility to track the full molecular motion in three dimensions.
Our results indicate that waveguiding semiconductor nanowires can detect single-molecules in 3D, allowing the investigation of molecular processes such as the diffusion of a single molecule or molecular motor movements with simple widefield microscopy without any modifications in experimental setup.
Image simulation pipeline (image)
1 D. Verardo et al, "Nanomaterials," 11(1), 227 (2021).
[2] J. Valderas-Gutiérrez, et al, ACS Appl. Nano Mater. 5, 9063–9071(2022)
[3] N. Anttu, 2024. doi: https://doi.org/10.48550/arXiv.2403.16537.
We give an overview of the work in AI within a joint research endeavour between the UKAEA and the STFC Hartree Centre, the Fusion Computing Lab, to accelerate discovery and design of complex facilities in magnetic-confinement fusion.
We illustrate how we are enhancing the "classical" real-time control of tokamak shots with AI techniques of emulation and reinforcement learning, both for the shape of the confined plasma and for the cooling of the exhaust fusion byproducts in the divertor region, as well as the fast prediction of disruptive instabilities from heterogeneous and gappy measurements.
We also present ongoing work to accelerate demanding simulations in plasma physics, to quantify turbulent transport in different regimes across the tokamak, from the core to the scrape-off-layer region, highlighting the importance of domain-aware AI.
The techniques developed in this collaboration also have applications to the efficient planning of experiment campaigns, and optimal design of complex facilities, in other areas of physics.
Many physical problems are governed by Laplace's equation, to which there exists a unique solution when boundary constraints are fully defined. However, given only partial knowledge or discrete observation along a boundary, statistical methods are required to construct uncertainty estimates and likelihoods for possible solutions.
Within magnetism, obtaining the magnetic field or the magnetic scalar potential, $\Delta \phi = 0$, from incomplete boundary information is of fundamental interest for instance in robotics navigation and geophysics when estimating the geomagnetic field on the Earth’s surface.
Here we propose and demonstrate new statistical methods for estimating the magnetic scalar potential from an incomplete boundary-value problem: (1) By using Gaussian Processes (GP) or (2) by an ensemble of physics-informed neural networks (PINNs). Using either we obtain statistical information for the complete boundary. This information is then used to efficiently construct the statistical solutions using a basis of harmonic functions for the magnetic scalar potential, such that this is guaranteed to be physically correct, i.e. obey Maxwell’s equations.
We show that the obtained uncertainty estimates using the above two methods compare well to the ones obtained from Hamiltonian Monte Carlo (HMC) simulations, which extensively investigate the target probability distribution for all valid magnetic scalar potential matching given knowledge.
Further, we qualitatively verify that the uncertainty estimates are correct, i.e., being confident where boundary values are given while being uncertain far away from observations points.
This allows for improved subsequent decision-making with the found statistical solutions to Laplace's equation, e.g., where to measure next when only limited amount of resources is available.