Fellows Call 1

Decision making methodology for waste management system design

Growing population, industrialization and capital are sufficient premises to forecast a substantial growth in waste generation in the near future. Waste treatment and management is directly responsible for 5% of total CO2 emitted worldwide. However, properly managing it has implications far wider. Fuels, heat and electricity can be obtained by properly valorizing waste, resulting in added value for several society sectors. This work pretends to design a decision-making methodology based on a comprehensive superstructure of waste conversion technologies, simultaneously connected with its logistics and supply chain.

Robust systems are designed taking uncertainty about costs and environmental impacts; this is extremely important in a world of fast pace and change – last April oil touched negative values and currently (November 2020) its value is on the lower end of the spectrum. It is thus vital to ensure that decisions made today remained valid, by properly analyzing trade-offs between different waste treatment configurations.

This work uses primarily modelling and optimization tools for generating a good set of solutions. Machine learning algorithms are being currently implemented to handle big sets of data and forecast decision-making trees.

Key words: Waste superstructure, CO2 reduction, Life Cycle Assessment, Mathematical modelling

Optimal training load estimation using personalized physical profile

Overtraining, an important risk factor for injuries in athletes, is the consequence of inappropriate and excessive training load (TL). Excessive TL is the result of the generalized methods used for TL estimation, which do not account for the individuality of athletes. This project aims to enable an objective assessment of the functional capacity, non-sport loads, and infield training quality of the athlete, and integrate them into a single framework to predict the optimal TL. This framework will utilize a personalized physical profile for each athlete, containing two main components: i) accurate capacity of based on instrumented field tests, ii) the training history containing the prior TL, and the quality of performed training sessions. By focusing on both, movement quality and overall TL, this physical profile can capture the effects of individual external TL on both physiological and neuromuscular adaptations. The profile can further serve as an input to a load estimation model, which can compute optimal training progression. The proposed load estimation framework constitutes a general approach to training load optimization. For its feasible implementation, two use cases are considered, training for endurance running competitions and reathletization process after injury. Running as a use case allows this work to build on the previous research in the lab while the second case is a first instance of the application of training load management to rehabilitation.

Key words: Wearable sensors, biomechanics, movement analysis, sports technology

Trajectory prediction of traffic participants

One of the main building blocks of an Autonomous Vehicle (AV) is “forecasting”. In this project, we are going to work on the problem of “sequence prediction” in vehicle domain. We aim to predict vehicles, cyclists, and pedestrians using available cues such as past observations, scene, and positions of other participants in the scene using deep learning methods. We are looking for domain-knowledge aware, interpretable and robust models. To go beyond pure imitation learning, we will inject our domain knowledge to the models. We further try to study the generalization of the models to communicate trust. Finally, the solutions should be robust to environment changes, failure cases (noisy detections), and adversarial attacks.

Technology Commercialization: Applications of Technology, their Feasibility and Ethical Boundaries

Entrepreneurs play a fundamental role in introducing (disruptive) technologies to markets. Because technologies generally have potential applications in multiple market domains, market opportunity identification has been seen as one of the most decisive decisions for the success of an emerging technology. Yet, while the importance of technologies as a source for wealth generation and economic progress is uncontested, the role of ethical boundaries in the commercialization of technologies has remained rather elusive so far. To address this research gap, the aim of this project is to develop an overarching model of how ethical considerations affect decisions regarding the application of technology and to delineate how different moral considerations play out in the technology commercialization process, particularly focusing on the trade-offs between financial feasibility / desirability and ethical constraints.

Meta-Learning for Versatile Artificial Intelligence

The goal of this PhD project is to enable machine learning models to grasp new concepts quickly with only a very small amount of data (few-shot). This setting is more common than one would assume, as small shifts between the learning and testing environment, such as lighting conditions for a camera or the position of a sensor, can lead traditional fixed models their performance to degrade easily. Adaptation to new or shifted concepts with very few new data is crucial in sustaining performance across such shifts. This domain is generally characterized as “few-shot learning” and is the main subject of this project.

Key words: Few-Shot Learning, Small Data, Machine Learning

SmartSwim: Novel smart swimming analysis system for exercise and training

As one of the three most popular sports in summer Olympic Games, swimming has always been attractive to study for sports scientists. Additionally, complicated nature of motions and variety of techniques made coaches need analysis systems. However accurate, existing analysis systems are too time-consuming and laborious for everyday use. With the advent of wearable sensors, inertial measurement units (IMU) in particular, motion analysis found the chance not only to study new aspects of motion, but also to cover in-field applications. Studies on swimming with IMUs mostly focused on extracting individual features, less aimed to use them to swimmers’ benefit. Hence, SmartSwim is going to offer useful details and evaluation of swimming in all swimming phases and styles, with a more practical approach. The main objective of this project is to develop a package of sensors and algorithms as an assistant for swimming coaches, broadening their view over the kinematic features of swimmers’ performance. It helps them with an objective evaluation of each swimmer individually and provides feedback for both coaches and swimmers. The main structure of this project is organized in three sections of feature detection, performance evaluation and feedback provision, each one dedicated to one step towards a novel swimming analysis system.

Key words: Swimming, Performance analysis, feedback, IMU

Neural mechanisms of olfactory repellence in the mosquito: Aedes aegypti.

The motivation of my project is to determine the neural mechanisms underlying mosquito odor avoidance behaviors and to use this knowledge to better understand the mechanisms underlying the behavioral change from attraction to repellence that is linked to odor concentration. Accomplishing these goals would constitute an important scientific advance toward the aim of reducing vector borne disease transmission. For this project, we have developed a high-throughput, long-term neural activity imaging and behavioral measurement technology capable of identifying how odorants recruit aversive olfactory circuits in Drosophila. We are now able to image neural activity in the Drosophila ventral nerve cord (VNC) (the insect equivalent of human spinal cord) during behavior. The VNC houses descending neurons that link the brain with motor circuits to control attraction and aversion to environmental odors. We will focus on understanding how compounds engage aversive neural circuits rather than simply aversive behaviors. This mechanistic focus can more effectively permit further chemical refinement to increase the potency of repellents and to give insight into how odorant concentration may transform an attractive behavior into an aversive one. Focusing on the neural substrates of aversion will allow us to understand and overcome the state-dependence and individual variability that typically plague purely behavioral screens. We will then transfer this technology to the mosquito: A. aegypti.

Key words: Drosophila, neural activity, long-term imaging

Supramolecular Thermoplastics for Automotive Applications

Supramolecular thermoplastics based on specific non-covalent interactions are a new class of polymers that have ductile features but can be conveniently processed from low viscosity melts. Up to date, these materials could not find their way into engineering applications, because they suffer either from too low melting temperatures or inferior mechanical properties.
This project, therefore, aims to develop novel supramolecular thermoplastics that combine both excellent ductile behavior and high softening temperatures. To this end, mixtures will be prepared from polymers end-functionalized with segments that can self-assemble into aggregates and match low-molecular-weight additives based on the same motif. The co-assembly of polymer end groups and additives will thus result in the formation of fibrillar aggregates. As a consequence, the materials’ melting temperature can be tuned by adjusting the concentration of self-assembling units irrespective of the molecular weight of the polymer segments. This approach significantly expands the scope of supramolecular thermoplastics by increasing the temperature range of operation to inaccessible temperatures, while allowing to independently control their mechanical properties. The materials can thus fulfill engineering criteria as imposed, for instance, by the automotive industry.

Keywords: Plastics, Recycling, Biodegradale materials, Supramolecular materials

Engineering novel quantum devices

The goal of this project is to understand, characterize, and engineer defect centers in large band-gap semiconductors. These systems have attracted increasingly attention thanks to their extraordinary optical properties and coherence time at room temperature, which enable them to be used for quantum computing, quantum information and sensing applications.

Despite these promising properties, many features of these systems are not fully explained and exploited. Among those, the optical absorption and emission mechanisms are not yet understood. In this project, we address this challenge by means of state of the art calculations based on non-equilibrium Green’s functions technique. We focus our effort on the negatively charged boron vacancy in 2D hexagonal boron nitride, which currently stands out among defect centers in 2D materials for its promise for quantum information and quantum sensing.

Furthermore, we perform high throughput first-principles simulations to search for novel defect centers in 2D materials whose properties overcome the limitations of those already studied.

A large share of the project, in collaboration with IBM, deals with the design of quantum algorithms for the current Noisy Intermediate-Scale Quantum (NISQ) computers, with the aim to accelerate the solution of condensed matter problems which scale exponentially on a classic supercomputer, and thus are not solvable but for very simple systems.
The goal of this project is to understand, characterize, and engineer defect centers in large band-gap semiconductors. These systems have attracted increasingly attention thanks to their extraordinary optical properties and coherence time at room temperature, which enable them to be used for quantum computing, quantum information and sensing applications.
Despite these promising properties, many features of these systems are not fully explained and exploited. Among those, the optical absorption and emission mechanisms are not yet understood. In this project, we address this challenge by means of state of the art calculations based on non-equilibrium Green’s functions technique. We focus our effort on the negatively charged boron vacancy in 2D hexagonal boron nitride, which currently stands out among defect centers in 2D materials for its promise for quantum information and quantum sensing.
Furthermore, we perform high throughput first-principles simulations to search for novel defect centers in 2D materials whose properties overcome the limitations of those already studied.
A large share of the project, in collaboration with IBM, deals with the design of quantum algorithms for the current Noisy Intermediate-Scale Quantum (NISQ) computers, with the aim to accelerate the solution of condensed matter problems which scale exponentially on a classic supercomputer, and thus are not solvable but for very simple systems.

Key words: Optical properties, quantum computing, 2D materials

Multi-agent Behavior Learning and Prediction for Autonomous Driving

While detecting and tracking important objects for Advanced Driver-Assistance Systems (ADAS) has experienced tremendous gains in performance in recent years thanks to deep learning, autonomous vehicles require a higher-level understanding of the dynamics of their environment to assess, operate in, and safely interact with that environment. The overall scope of this project is to perceive, learn, and predict the behavior of multiple agents of different types (pedestrians, vehicles, cyclists, etc.) from a large number of dynamic scenarios and a set of detections and tracks, to enable better Autonomous Driving systems.
The aim is to learn a robust representation that is useful in all the aforementioned tasks while it is invariant to a set of nuisances. We want to study the task of discovering that representation with a new paradigm of learning by visual prediction. To do that, we propose new deep generative frameworks as visual predictors in a self-supervised learning (using the signal itself as the free data) setting. This can enable us to leverage massive free unlabeled data needed for training our model.

Key words: Visual prediction, Self-supervised learning, Deep generative models

Collision resilient drones for long-range operations

Aerial robots have increasingly been used in a significant number of professional and non-professional applications today ranging from entertainment to search and rescue missions. However, flying robots suffer from several limitations, especially when they have to negotiate heavily cluttered and chaotic environments. Most drone configurations are susceptible to damage from external collisions because of their rigid configurations, their limited protection, and exposed moving components such as their propellers or their control surfaces. Although drones that are collision resilient exist both in the literature and on the market, collision resilient strategies that allow drones to continue flying even after a collision are limited to rotorcrafts with limited range and operational endurance. The purpose of this thesis is to develop the next generation of collision resilient, autonomous, flying robots. The focus will be on inspection, monitoring, and reconnaissance operations for emergency scenarios. The flying platform proposed to be developed will be able to operate while hovering in cluttered environments, to successfully manage collisions, and be able to adapt its configuration to travel for long distances in order to ensure human operators’ safety while reaching possible incident areas in time.

Key words: Drones, Shapeshifting, Safe, Resilient

Personalized Detection and Prediction of Epileptic Seizures in Real Time with Wearables

Epilepsy is one of the most common chronic diseases, affecting around 50 million people worldwide. Despite advances in treatment, one-third of patients still suffer from refractory epilepsy, experiencing uncontrolled or frequent seizures. Hence, we propose a long-term real-time wearable monitoring system that can potentially improve the quality of life for such patients and reduce the mortality rate due to seizure-related accidents and sudden death in epilepsy. There are two essential requirements for modern wearable health monitors: 1) early and accurate detection of pathological conditions and 2) monitoring the patients in real-time and on a long-term basis. The design of accurate and energy-efficient wearable wireless body sensor networks involves several challenges in
terms of detection performance and system battery lifetime. To tackle these problems, this research project proposal brings as key innovation topics the following ones: 1) multimodal biosignal processing for epileptic seizures detection, combining the information in (4 electrodes hidden in the glasses) EEG and (2 electrodes hidden in the shirt) ECG signals, to reach a medically-acceptable accuracy level; 2) the new event-driven energy-management paradigm, inspired by the brain’s remarkable efficiency, to enable the high energy-efficiency required for real-time and long-term monitoring and to compensate for the complexity of the multimodal biosignal processing.

Key words: Epilepsy, EEG, Wearables, Edge Computing

Quantum LiDAR and non-line of sight imaging with SPAD camera

My research include two topics, the scientific research on quantum LiDAR (light detection and ranging) and the engineering research on non-line of sight imaging. Both topics are based on a SPAD (single-photon avalanche diode) camera, which will be designed by myself to achieve functionality and performance goals.
Imaging with quantum states of light can use spatial correlations between photon pairs to overcome fundamental limits of classical imaging, like performance enhancement in the presence of noise. Most of the research used electron multiplying CCD (EMCCD) technology. Compared to EMCCD, SPAD technology has the advantage of increased acquisition speed and time-resolved information. The goal of this scientific research is to combine quantum imaging and time-of-flight to realize a true quantum LiDAR.

Non-line of sight imaging (NLOS) allows objects to be observed when partially or fully occluded from direct view, by analyzing indirect diffuse reflections off a secondary relay surface. Example applications include search and rescue, ADAS, and autonomous driving. In recent years, SPAD cameras with time-resolved information have led to advances in this application. The goal of this engineering research is to pave the way towards commercialization by building a compact and reliable NLOS prototype.

Key words: SPAD, Quantum Imaging, NLOS


Funded by
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 754354.