Fellows

Materials discovery involves the search for new materials with unique properties and applications. Traditional methods rely on time-consuming trial-and-error approaches, where researchers synthesize and test numerous materials to find those with desired properties. Machine learning (ML), particularly, can significantly expedite this process by autonomously learning patterns from data, offering faster results. However, machine learning in materials discovery faces several challenges. Firstly, datasets are often small and lack a unified representation, limiting the models’ ability to generalize. Secondly, material data comes in various formats, such as microscopic images, tables, and chemical structures, making data integration a complex task. Moreover, there is an abundance of unlabeled data that could be leveraged in the ML process, but existing models fail to do so. This project aims to maximize the use of available data sources to develop robust and reliable models. It will create a unified embedding space for materials, allowing integration of various data types. A novel semi-supervised active learning strategy will incorporate knowledge from unlabeled data while identifying remaining model vulnerabilities and suggesting mitigating experiments. This approach will accelerate research in materials science by simplifying method comparisons and render property prediction models more accurate, robust, reliable, and generalizable – essential traits for materials discovery.

The ability to recognize objects beyond the training taxonomy is essential for computer vision models in safety-critical applications such as Unmanned Surface Vehicles (USVs). USVs are exposed to a wide variety of possible obstacles which is impossible to cover by collecting and labeling a dataset. Recent advances in multimodal vision and language learning have enabled training models that are capable of recognition on open vocabulary. This research proposal aims to investigate open-vocabulary panoptic segmentation within the maritime domain. This domain presents unique challenges due to limited labeled examples and a constrained training taxonomy in maritime datasets. To address these challenges, we intend to explore semi and weakly supervised learning techniques for open-vocabulary models. Moreover, USVs cannot depend on cloud processing due to the high latency and unreliable nature of mobile networks. As a result, all control critical computations must be conducted using embedded on-vehicle hardware. This prevents the deployment of current open-vocabulary segmentation models due to computational inefficiency. Therefore, this research proposal aims to investigate embedded-ready architectures for open-vocabulary panoptic segmentation.

Environmental and socio-economic challenges have been compounding in the last decades and present significant problems for the long-term future of the modern globalized world. There is a fundamental scientific and practical necessity to adequately understand human-environmental dynamics at multiple scales. The HESTIA project proposes a systematic approach to apply automated scientific discovery frameworks, especially on identifying systems of differential equations, to several challenging problems, such as modelling long-term societal developments and the possibility of collapse, to quantifying large scale macroeconomic patterns and environmental changes. The main challenges for the project come from the inherent complexity of socio-environmental systems (with non-linearities, feedback loops and cross-scale interactions), from data and modelling limitations (integrating and managing diverse datasets, bridging interdisciplinary gaps) and from model complexity, which posses difficulties for interpretation, transparency and communication. HESTIA aims to use automated equation discovery frameworks to allow model development which incorporates expert knowledge and uncovers causal structure. Furthermore, the frameworks can deal with coarse data and provide an interpretable set of equations and output, which greatly aid communication, both scientifically and in policy.

The primary goal of this research project is to develop and implement a novel class of methods based on evolutionary computation (EC) that are suitable for training reinforcement learning (RL) agents in uncertain and dynamic environments in a few-shot and model-agnostic (without strong distributional assumptions) manner. To achieve this goal, special attention will be paid to a novel class of EC algorithms based on the ideas of divergent search and open-endedness, which particularly includes quality-diversity approaches, capable of generating a wide range of mutually diverse yet high-performing solutions. Such diversity in the population is hypothesized to facilitate tracking moving optima in non-stationary environments, all while achieving significant sample efficiency. The proposed research will impact the state-of-the-art in the burgeoning field of evolutionary RL by combining EC with few-shot online RL. Moreover, it will touch upon multiple topics in the related areas of stochastic multi-armed bandits and risk-sensitive RL. A significant focus will be put on designing and investigating exploration strategies and policy regularization techniques suitable for this context. Moreover, the project has a clear multidisciplinary nature, with two secondments aimed at applying the developed methods to domains in the area of climate research, particularly flash floods over the mainland, and remote sensing in the context of the study of aerosols.

Firstly, the proposal is to explore the application of Machine Learning (ML) especially beyond supervised learning in enhancing the security of wireless communications within the terahertz band for future scientific applications. Terahertz frequencies offer immense bandwidth and potential for high-speed data transfer, making them pivotal for future communication systems. However, they pose unique security challenges, such as vulnerability to eavesdropping. This research aims to develop ML-driven techniques for secure data transmission in this challenging spectral range, ensuring the confidentiality of information. Secondly, the research contributes to advancing the field of AI and wireless communications for scientific applications, which is a vital interdisciplinary area under SMASH fellowship. By harnessing AI's capabilities to optimize signal processing, beamforming, and encryption in the terahertz band, this research will address current limitations and open new possibilities for communication technologies.

The Galactic Center Excess (GCE), a GeV gamma-ray excess observed by the Fermi Large Area Telescope (Fermi LAT), has attracted significant interest as a potential dark matter (DM) signature. However, determining the GCE's true origin is hindered by complex astrophysical backgrounds and uncertainties. In this research project, I propose an innovative approach to deciphering the GCE by utilizing advanced statistical methods and machine learning techniques. My strategy encompasses modeling DM signals, investigating astrophysical backgrounds that impact DM searches, and analyzing gamma-ray data from both the Cherenkov Telescope Array (CTA) and Fermi LAT. I will employ simulation-based inference (SBI) to achieve the necessary flexibility in gamma-ray emission models and to address observational systematic uncertainties often neglected in current literature. A key component of my project is the development of an SBI gamma-ray analysis pipeline also suited for CTA datasets, maximizing the potential of current and next-generation instruments in gamma-ray astronomy. By elucidating the GCE and exploring non-thermal astrophysical processes in the Galactic center at GeV and TeV energies, this project aims to advance our understanding of DM and high-energy astrophysics. Simultaneously, it offers a novel approach to gamma-ray data analysis through the application of SBI.

The need for highly accurate large-scale simulation of particle collisions at the Large Hadron Collider motivates the development of new simulation techniques. Generative deep learning algorithms have been successfully used in various scientific fields, including High Energy Physics. Generative adversarial networks are already applied in the simulation of ATLAS calorimeters. The aim of the project is to translate the idea on silicon tracking detectors in order to cope with the increasing number of sensitive detectors in the new ATLAS Inner Tracker built for High-Luminosity LHC.

After twenty-five years from the discovery of the late-times accelerated expansion of the universe, the Standard Model of cosmology (ΛCDM) remains unsatisfactory from both the theoretical and observational points of view. Not only the nature of Dark Energy (DE) and Dark Matter (DM) remains a mystery, but current cosmological tensions may indicate the presence of new physics beyond that provided by the standard ΛCDM paradigm. In the next decade, Large Scale Structure (LSS) experiments will play a crucial role to understand the properties of the exotic components and to shed light on their nature. Extracting most of the cosmological information contained in the data that the next generation of LSS experiments will provide us is the goal of an entire community. To get the most from these data, it is crucial to rely on fast and robust tools to estimate LSS observables in alternative cosmologies. The aim of this proposal is to pursue this road, making use of machine learning techniques to develop fast and accurate tools for DE and Modified Gravity (MG) models. My plan is to (i) build an emulator for the necessary Power Spectra (PS) in DE/MG scenarios; (ii) integrate it in existing pipelines to get LSS PS; and (iii) use this machinery to get the most up-to-date constraints with existing data and forecast the sensitivity that forthcoming LSS experiments, with particular attention to Rubin Observatory Legacy Survey of Space and Time (LSST), will have on cosmological parameters.

The nature of Dark Matter (DM) remains one of the most important open problems in fundamental physics. Despite a plethora of theoretically well-motivated candidate particles, definitive evidence of DM has not been discovered in a wide-ranging program of experimental searches. The production of DM and other particles beyond the Standard Model (SM) can potentially leave signatures of new physics (NP) at the Large Hadron Collider (LHC) and the interactions of DM with nuclei can be probed by direct detection (DD) experiments. This proposal focuses on LHC searches relevant for charged mediator models, in which DM-SM interactions are mediated by an additional NP particle with SM charges. The primary research objective is to develop machine learning (ML) techniques for analyses of LHC data in searches for charged mediator models which are difficult to probe with cut-and-count approaches. Both fully supervised and unsupervised ML algorithms will be implemented to discriminate NP signal from SM background. Several of the ML techniques which will be developed for LHC analyses will also be implemented for DD searches using mineral detectors, which have been proposed as an alternative to conventional DD experiments. Rather than relying on increasing large instrumented volumes of target nuclei monitored for several years in search of rare DM interactions, mineral detectors record and retain damage tracks caused by nuclear recoils from DM interactions over geological timescales.

The Vera Rubin Observatory (LSST) will collect an unprecedented quantity of data on variable astronomical objects. One of LSST's goals is to detect potentially new transient phenomena and to obtain a more complete view of the various types of transient (sub)populations such as SuperNovae and Tidal Disruption Events. To address the mission's objectives, tools that can perform automated class discovery and topographical analysis of the distribution of the light curves will be crucial. One key challenge to do so is to obtain informative, fixed-size representations (encoding) of the time series. We propose to explore and vet the abilities of Recurrent Neural Networks to encode the relevant information in the representation space. Through a careful analysis on toy and simulated datasets, we aim at getting a better understanding of the properties of RNNs as dimensionality reduction methods for LSST's time series, and the impact of using additional constraints during training. We will then evaluate the potential of Density-Peak Based Clustering for automated class discovery. We also propose to address the problem of similarity and metric learning in the context of LSST's time series with a contrastive learning approach: our goal is to develop a method that can find similar examples in a large unsupervised set, using only a few positively labelled ones. Such approaches will potentially aid in increasing the detection rate of some objects (e.g. TDEs from lower-mass black holes).

Clusters of galaxies are the largest, most massive virialized structures in the Universe, and highly dominated by dark matter (DM). According to the most accepted scenario, they should shine brightly in gamma-rays. This gamma-ray diffuse emission is expected from the decay of the neutral pions created from the hadronic interactions of cosmic ray protons; but also, from the annihilation and/or decay of DM particles. Despite the efforts in the last decades searching for this diffuse gamma-ray emission, such signal remained elusive. However, recent works strongly suggest that we may finally be on the verge of the first conclusive detection. My SMASH research project aims to investigate the challenges associated with the detection of diffuse gamma-rays from galaxy clusters and expand our understanding on the different processes that originate this emission. To achieve this ambitious goal, I will exploit at its maximum the data from existing gamma-ray telescopes (Fermi-LAT) and explore the capabilities of the future Cherenkov Telescope Array Observatory, using a Convolutional Neural Network based pipeline especially trained for the identification of faint sources (as clusters of galaxies) in complex environments, involving extended and overlapping sources. The use of this novel pipeline based on Machine Learning techniques will serve as benchmark for future CTA analyses and due to its inherent adaptability, will also aid solving other gamma-ray puzzles related to faint sources.

Axion-like particles (ALPs) are promising candidates from theories beyond the Standard Model, possibly linked to dark matter. When subjected to external magnetic fields, ALPs can convert to photons and vice versa, rendering them observable.  This project delves into the study of these interactions within the astrophysical realm, focusing on the gamma-ray spectra of the blazars. Blazars are the relativistic plasma jets emitted from the intense conversion of gravitational energy in the vicinities of super-massive black holes and rank among the brightest gamma-ray sources in the sky. The ALP-photon mixing distorts the gamma-ray blazar spectra with measurable effects, albeit tiny. The description of blazar jets varies per target and demands multi-wavelength data but a global study of the jet in the ALP scenario is still missing. We first aim to target this issue by performing a structural study of blazar jets with open-source modeling tools, including polarization measurement. Additionally, we harness the power of the supervised Machine Learning (ML) approach to investigate the interplay between photons and ALPs, leveraging gamma-ray data from the Cherenkov Telescope Array (CTA) and the Large Size Telescope (LST). By training the ML network with a large set of simulated data from CTA and LST for a diverse set of magnetic fields and coupling constants, we seek to yield robust constraints on ALP-photon interactions and make substantial advancements in this field.

This project is devoted to developing the most detailed comprehensive numerical framework to date, combining magnetohydrodynamics (MHD) and Monte-Carlo simulations to derive the multi-wavelength (MWL) and multi-messenger spectra from the magnetized environment of galaxy clusters by comparing the energy gain of accelerated particles (i.e., protons, heavy ions, and electrons) with their radiation losses. Special attention will be given to Perseus-like clusters hosting active galactic nuclei (AGNs). This work is an extension of our previous research. To probe the intracluster medium (ICM) across redshifts, we will develop MHD simulations including AGN-feedback and star-formation. To study the cosmic ray (CR) transport in the ICM considering all relevant interaction processes, the multi-dimensional Monte-Carlo simulations are well suited. To ensure the robustness and timely production of research products, machine learning and data analysis techniques will be used, extensively. This work can appropriately interpret the observed emissions from clusters and provide source distribution constraints for various observation facilities such as CTA, HAWC, LHAASO, IceCube-Gen2, TA, and Pierre Auger Observatory, etc. Furthermore, this numerical method can be applied to study the multi-frequency emission from various types of astrophysical sources and such investigations with an unprecedented level of detail can address several open questions in multi-messenger astronomy and cosmology.

Weather forecasts are subject to uncertainties originating from various factors, including initial condition uncertainties, model imperfections, and numerical errors. Data assimilation (DA) is a critical process in Numerical Weather Prediction (NWP) that fuses observations with model simulations to refine the initial atmospheric state for accurate forecasts. DA is computationally intensive, and traditional methods require reducing the computational space using linear methods. Neural Networks (NNs) could solve this challenge by enabling nonlinear reductions. Nevertheless, their application in NWP DA has remained limited so far. This proposal builds upon recent Neural Network DA (NNDA) advancements, particularly those using autoencoder NNs for dimension reduction. A novel approach leverages a Variational Autoencoder (VAE) convolutional NN, applying the 3D-Var cost function in the latent space, drastically reducing the background error covariance matrix dimensions. This innovation simplifies the inversion process, maintaining Gaussian properties in the latent space. This project focuses on constructing an encoder-decoder observation operator that models radar data from atmospheric model fields, allowing comparisons with observed data. The proposed approach aims to advance operational weather prediction by enabling more efficient DA and providing near-continuous estimates of atmospheric states, benefiting not only meteorology but also a wide range of scientific disciplines.

Exposure to ambient particulate matter is responsible for adverse health effects, while aerosol and climate feedback mechanisms are still obscured by large uncertainties, related to the vast temporal and spatial variability of ambient concentrations and aerosol sources. Source apportionment (SA) is an important research field, providing valuable information on the presence and variability of sources of particulate matter, acting as an essential tool for informed decision making aiming at abatement of atmospheric pollution through regulating specific polluting activities. SA of organic aerosol (OA) is commonly performed via positive matrix factorization (PMF) on OA mass spectra, obtained with aerosol mass spectrometric techniques. PMF suffers from certain caveats, i.e., non-unique solutions, significant level of subjectivity and modeler bias, bias in the source mass spectrum characterization, inability to follow varying number of sources present and demands significant computational time. In this project a novel method implementing a two-step machine learning OA SA technique will be developed, tackling the problems listed above. The technique is trained on an extensive library of deconvolved and source specific OA mass spectra and is evaluated on four distinct long-term ambient OA datasets, characteristic of both urban background and traffic conditions, and three datasets obtained in cities heavily polluted by sources like traffic, biomass burning and trash incineration.

Aerosolized Particulate Matter effects on human health and climate depend on its composition, hence proper classification and determination of sources are necessary for effective abatement. This is performed by source apportionment (SA), the process of identification of the sources of air pollution, usually conducted through receptor modelling (RM). The usual approach is decomposing the input measurements into the product of time evolution (time series) and chemical composition patterns (profiles) of pollutant sources. However, RMs struggle to provide well-separated profiles, leading in turn, to mixed time series contributions. To address this issue, we propose the integration of data science tools, including sparsity approaches, into novel and state-of-the-art RMs to aim for a better source differentiation. Performance assessment of RMs is usually conducted through synthetic datasets, but current approaches contain design biases or are of limited environmental reproducibility. In order to evaluate more accurately the RM performance and validate the RMs, our approach consists on creating new methodologies for synthetic dataset creation through machine learning from real-world measurements and modelling outcomes. In such way, we will reach the ultimate objective of this project, which is to upgrade SA techniques guaranteeing further source differentiation to improve the monitoring of atmospheric pollution composition and focused abatement of air pollution and climate change.

The MEGLICA project aims to tackle a series of problems in the current state of the art of the aerosol science, and more in particular in the topic of the warming effect of absorbing aerosols on climate. In fact, the contribution of light absorbing aerosols (LAA), and in particular BC, the highest aerosol contributor to aerosol absorption, and one of the highest contributors overall, has presented 4-fold decrease on its effective radiative forcing on climate in the newest IPCC report (2021) in comparison with the previous one (2013). The large uncertainties still present prevent from correctly constraining their effect on climate. The MEGLICA proposal aims to reduce these uncertainties by accomplishing several steps for which it will use Machine learning as a tool to help with the analysis: first, we propose a novel calibration technique using the potential reference methodology PTI for the measurement of the aerosol absorption and BC concentration, then we intend to improve the LAA global spatial heterogeneity in 3-dimensions and determine the atmospheric heating rate due to their presence on the atmosphere, and finally, to use these improvements to model the LAA radiative effect on the climate.

Intensive and meteorologically induced high frequency oscillations (HFOs) of sea level occur occasionally in several Mediterranean locations. In special cases they reach destructive heights known as meteotsunamis. Despite the wide knowledge of physical dynamics of these phenomena, available forecasting systems are unreliable and numerically expensive. The goal of the HF-SCANNER project is to create a reliable and fast forecasting system of the Mediterranean HFOs and meteotsunamis, using machine learning methodologies and simulated (ECMWF) and observational (sea level and air pressure) data. The project has three major steps. The first one is to create pre-trained supervised and self-supervised foundation ensemble deep learning (DL) models using long observational series available at the two locations. The second is to transfer these models, using transfer learning, to locations with less observational data. The final step is to create a DL-based forecasting system to predict an amplitude class of upcoming HFOs and meteotsunamis in several Mediterranean harbors. The proposed project will address HFO modeling challenges in an original and novel way by applying advanced DL methodologies and by using the longest sea-level series available in the Mediterranean, as well as a 50-member ensemble of full 3D simulated atmospheric fields. The result of the HF-SCANNER project will solve a scientific problem but also it will address existing societal hazard risk.

The integration of machine learning (ML) with advanced biomarker discovery techniques offers new opportunities for pathology, particularly in personalized medicine. This research will focus on using nanobodies—small, stable, and highly specific single-domain antibodies derived from camelids—as versatile tools for advancing biomarker research. We plan to utilize ML to explore a diverse, naïve nanobody library, with the aim of mapping the "hidden human epitopome"—an unexplored landscape of potential antigens and epitopes. To achieve this, we will combine high-throughput sequencing and state-of-the-art ML-driven 3D structural modeling with deep-learning-based molecular docking. This methodology will allow us to systematically predict and validate nanobody interactions with a wide range of human protein structures, ultimately creating a comprehensive catalog of nanobody-binder pairs. We will validate this approach through biophysical and biochemical assays, complemented by structural biology techniques. The goal of this research is to uncover novel biomarker-epitope interactions, enhance our understanding of disease mechanisms, and contribute to the development of new diagnostic tools and personalized therapies. By bridging advanced computational methods with the biological potential of nanobodies, we aim to redefine biomarker discovery and provide a robust platform for future therapeutic interventions.