Results 2021-1 Industrial Fellowships Call

The FNR is pleased to communicate that 8 of 12 Industrial Fellowship projects have been selected for funding in the 2021-1 Call, representing an FNR commitment of 1.54 MEUR. 

The aim of the Industrial Fellowships programme is to foster the cooperation between Luxembourg based companies active in R&D and public research institutions in Luxembourg and/or abroad. The scheme awards PhD and Postdoc grants to researchers who carry out their PhD and/or postdoc training in collaboration with a company in Luxembourg. The scheme is open to all scientific domains, and do all researchers, regardless of their nationality. Collaborating companies must have a presence in Luxembourg.

Go to Industrial Fellowships programme page

Funded Industrial Fellowships projects

PhD – Domain – Information and Communication Technologies

Abstract

Orbiting satellites and spacecrafts around the Earth offer a wide range of applications that affect our daily lives. These applications include weather forecasting, high speed internet provisioning, reliable telecommunication channels, television broadcasting and many more. As time passes, more and more orbiting objects have appeared such as satellites, nanosatellites, space debris, comets and asteroids. Within this context, there has been a huge increase in the number of space missions involving close proximity with an orbiting object, for example when docking to a satellite or repairing it. To ensure the safety of these missions, satellites generally rely on calculating the position and orientation, known as pose, of the closing object. The goal of OBSERVE is to develop a novel Artificial Intelligence-based pose prediction solution which is capable of handling a range of different orbiting objects. The developed AI solution will benefit from recent advances in pose prediction research using images captured by a regular colored camera. Furthermore, OBSERVE also aims to ensure that the AI solution is suitable for space. All of the research and developments conducted within this project are done in close collaboration with a space technology start-up based in both United Kingdom and Luxembourg.

Abstract

Goodyear (GY) provides tires and tire-related mobility services to its global customer base. The operations of global enterprise are very complex and large scale. GY makes tens of millions of tires each year, using several factories across the world, while each tire is composed of several components, and each component requires specialized machines.

Management of such systems is intrinsically hard, as they involve a network of processes with complex interactions. The number of interactions does not allow to simply analyze the system using statistics or classical operations research methods. However, thanks to the big data and simulation possibilities, it is nowadays possible to gather all the data, feed it to the simulation and analyze the results. It is also possible to change the configuration and run the simulation again. For a factory example, we could be interested in the number of tires that would be produced if we changed the production sequence of components.

The immediate question that follows this observation, is how to find the best configurations for the analyzed systems. This is an optimization task. Intuitively, but also mathematically, optimization is easier for systems with fewer possible setup variables (called decision variables), and when the output of the system is linked to the decision variables by direct and simple relationships. Unfortunately, optimization of operating systems of an enterprise involves thousands of decision variables, with non-trivial relationships. While such tasks may seem daunting for a human operator, there exist important advances in the field of Artificial Intelligence and Computer Science to solve such problems. They are algorithms that enable simulation-based optimization, that is optimization that uses simulation in the process of search for the best configuration.

In this project, we focus on two families of such algorithms. Metaheuristics are search methods that explore the search space by introducing perturbations to proposed configurations and subsequently evaluate the quality of new configurations by running the simulator. This process is randomized, and is aimed at balancing exploration of new configurations, and exploitation of the already gathered knowledge. Deep Reinforcement Learning (DRL) is a method in which the algorithm learns the best behavior running inside the simulation, by continuously making decisions on how to configure the system and observing outcomes. Currently, GY uses metaheuristics for simulation-based optimization.

Metaheuristics can effectively solve small-scale problems, however, they suffer from two practical drawbacks. First, the effective use of metaheuristics requires an expert to tune it to the specifics of the problem. As such, they are not generalizable, which results in additional time to develop a well-performing metaheuristic. The other gap is the understandability of the configuration and the rationale behind its performance, which increases the need for validation and decreases the trust of the final users.

The goal of this research project is therefore to close those gaps by dedicated hybridization of metaheuristics with DRL. We aim to introduce learning capabilities into the search procedure, such that promising regions are identified quicker, and this process is automated. This would close the scalability and generalizability gaps. We also intend to allow algorithms to redefine the way a configuration is described, and by measuring the size of the description and automatically decreasing it, we intend to improve the understandability. The results of this project will therefore directly improve the performance of GY operations, but they will also bring fundamental advances to the area of optimization, and they will be transferable to other industries and services.

Abstract

With the advancement of FinTech, in parallel with the Luxembourg FinTech initiative and a new paradigm of personal data regulations, banks in the world are getting more and more interested in Federated Learning (FL) technology and its use cases in KYC (Know Your Customer), KYT (Know Your Transaction), and AML (Anti-Money Laundering) systems.

With the collaboration of the Banque et Caisse d’épargne de l’État (Spuerkeess or BCEE), the objective of this project is to focus on data privacy-preserving using FL technologies for retail banking. More accurately, within the field of data-driven KYC/KYT using graph neural networks to build a 360-degree view of customers with the goal of better customer satisfaction, long-standing relationships and prevent customers from fraudulent transactions through their activities and interactions.

The subject will be targeted on graphs for machine learning (ML) using Federated Learning in self-supervised framework to overcome the lack of labels in the industry and avoid the cost of data annotation.The project focuses on improving the current state-of-the-art in Heterogeneous Attributed Networks (HANs) to provide contextual inference about customers in KYC and KYT using Graph Neural Networks (GNN) Embedding. On the other hand, we need to preserve the customer’s privacy following new regulations set by CNPD (Commission Nationale pour la Protection des Données) of Luxembourg. With this in mind, we benchmark, study the feasibility and the efficiency of Federated Learning in real-world use cases in a virtual environment (VE) integration, VE Integration provides protection of VMs in virtual BCEE server environments to use BCEE data, to compare different FL schemas in terms of their data privacy-preserving properties to match the Luxembourg regulations and BCEE needs.

Additionally, the project will also focus on the development of various GNN architectures to maximize the use of the graph data structure and attributes. Our objective will be to extend the use graph Embeddings to investigate suspicious, anomalous, and malicious behavior patterns in banking transactions, by developing hybrid spatial-spectral GNN architecture to fully exploit the insights simultaneously from both the spectral and spatial graph domains.

Noted that some anomaly detection in signal processing (signals or time series) requires the knowledge in both time domain and frequency domain. In this way, Federated learning could be extended to reduce the false positive rate using GNNs in Transaction Monitoring System (TMS) as part of AML system, where the TMS will typically use information from KYC and KYT to interrogate every transaction completed by a customer before authorize the execution.

To conclude, our innovative research for graph data analysis in retail banking will promote research of combined algorithms between convolution operations and shallow neural networks for spatial and spectral graph representations for deeper embedding analysis for identification and prediction retail banking downstream tasks. In the same scope, our innovative approach extends our work on data-driven KYC/KYT using Heterogeneous Attributed Networks GNNs in Federated Learning, this will promote research on collaborative learning with privacy-preserving and ML for financial institutions. The tools and algorithms developed during this project will be in direct implementation within our industrial partner BCEE.

Abstract

With the fast-pace advancements in robotics automation across companies worldwide, researchers focused on the idea of pulling robots out of their cages and share their workspace with humans, a concept called collaborative robotics. While traditional industrial robots have a fixed program, collaborative robots (cobots) are meant to be in a dynamic environment, in terms of interacting with humans and task variety.

Transforming into cobots is a crucial need for many companies like our industrial partner, Rotarex, as production scheme moves from mass production to mass customization. In mass customization, there are several different products variants with slightly different components and assembly procedures, leading to a large number of possible products. Currently, the cobot at Rotarex facilities is not operating with enough accuracy and robustness, and also not in a flexible mode suitable for mass customization. This is mostly because programming a collaborative task and also making modifications for adapting to a new task is a time-consuming and costly process and mainly requires the presence of a robotic expert.

Researchers have reached promising results about how Machine Learning (ML) approaches would improve this issue, where robot can learn the collaborative task from human demonstrations while learning the uncertainties of the environment to increase robustness and precision. However, past works either do not support the required task complexity in a real industrial setting, or rely on a robotic expert to implement in industry. With this in mind, this project’s aim is to support the engineering of automated but highly flexible assembly processes using cobots in a fast and cost-efficient way by exploiting different ML approaches in a suitable combination.

The final goal is to provide a framework that allows Rotarex process engineers to integrate cobots in the assembly process without extensive modelling and costly support from robotic experts. Our contribution will first include improving state-of-the-art ML methods to learn more sophisticated tasks from human demonstration. Then, we will extend our approach to integrate Human-Robot-Collaboration (HRC) based on task needs.

To avoid complex modelling, we aim at pre-training the robot about the task model by human demonstration, and then improve the model and overall performance in real time and during robot operation, similar to the way humans acquire skills by trial and error, so that the robot adapts to more precise and accurate procedures. Finally, our main contribution is to take the overall assembly into consideration (and not a single task) and provide an engineering toolbox based on our findings, which will be directly applicable by process engineers as a quick and efficient adaptation tool for cobots in Rotarex.

We will provide proof of concept by experimenting, measuring, and validating the performance of our framework in real applications, and finally publish our scientific outcomes in the form of a PhD thesis and open-access publications in peer-reviewed international journals in compliance with FNR’s Plan S.

PhD – Domain Physics, Engineering & Material Sciences

Abstract

Currently the linear economy is predominant is most sectors and the construction sector is not an exception. The linear economy which is largely based on the principle “take-make-waste”, it assumes that natural resources are available, abundant, easy to source and cheap to dispose of. However, this is not sustainable as the world is moving towards planetary boundaries, this model is causing resources depletion, and excessive greenhouse gas emissions and waste. Consequently, a transition of the current linear model to sustainable models capable to improve the efficiency and effectiveness of resources use is urgently needed.

The circular economy approach in the built environment promotes the optimization in the use of materials, goods and components in order to decrease waste generation and resources depletion to the largest extent. This is not contrary to the definition of sustainability; indeed, it can be defined as a way to reach sustainable development goals, based on reduction of waste by reducing, reusing, recycling and recovering materials in production, construction and during use of buildings. However, there is still lack of circular construction alternatives and research to promote the transition to the circular economy model.

The construction’s sector transition from a linear to a circular economy is in line with the European Green Deal which aims to foster the efficient use of resources by moving to a clean circular economy to be climate neutral by 2050. In addition, in 2020 Luxembourg established Sustainable and Responsible Development as one of its national research priorities in which implementation of circular economy is included to achieve transitioning to a sustainable community. Therefore, in order to reduce resources depletion, carbon emissions and construction and demolition waste (CDW) caused by the construction sector, novel construction practices targeting design for deconstruction and reuse of components are urgently required. In this context, the Prefa-SeTi (Steel-Timber Composite Beams) project, which is a collaboration between the ArcelorMittal Chair of Steel Construction at the University of Luxembourg and the PREFALUX (Luxembourg), surges for minimising the environmental impacts of the construction sector. It aims to contribute to decarbonization by fostering a circular economy with reusable LEGO®-like and sustainable steel-timber structures. Due to lack of research and codes for design and construction targeting steel-timber composite construction it is currently very difficult to implement such hybrid systems in practice.

This research aims to provide a basis for the use of novel steel-timber hybrid structural solutions. To achieve this, a new demountable steel composite flooring system will be investigated: engineered timber panels will be used as reusable floor elements connected to steel beams. Innovative shear connections between the steel beam and the timber elements will be developed and tested. The basic idea of SeTi is to enhance the disassembly, reuse and inter-exchange of structural steel-timber composite flooring elements with standardised sizes.

Like LEGO®, the structural elements are connected together in such a way that they can be easily disassembled and reused. Therefore, novel standardised and dismountable connections will be developed and implemented instead of one-way screws and welded connectors. This structural system will be thoroughly investigated by means of laboratory tests and advanced numerical simulations. The information obtained in the tests and the simulations will be the basis to develop equations to predict the behaviour of the flooring system.

Abstract

Today, the space industry shows a clear trend to enable challenging applications and decrease mission costs by using robotic manipulators that enable more autonomous operations. In order for such systems to work autonomously, it has to ‘see’ and ‘understand’ the environment through the camera(s). The use of cameras for visual perception in this context is associated with the low sensor costs combined with dense environment information. This allows flexible use of the same perception system/manipulator for various complex scenarios such as satellite servicing, in space assembly, and resource utilization.

The prospect of building such promising systems is becoming increasingly possible with the latest advancements in computer vision, robotics, and machine learning, coupled with improving on-board computers for commercial spacecraft. However, the development of perception systems for autonomous space-borne manipulators is challenging as high effort is required to ensure reliability while managing complexity and flexibility. Additionally, drastic lighting conditions in space and limitations of the on-board computer requires algorithms to be specially optimized for use in space. To tackle these challenges, in this project, we propose the development of a novel vision pipeline to enable dynamic grasping, a fundamental robot skill, in space.

We propose to develop the vision pipeline that is autonomy-ready and computationally efficient, that is suitable for operations on-board a spacecraft. Further, we assess robustness and simulation to reality transfer of this vision pipeline in mission relevant test scenarios to close the gap and demonstrate practical relevance of the developed vision pipeline.

Postdoc – Domain Materials, Physics & Engineering

Abstract

Designing tires is at the core of Goodyear’s business as a tire manufacturer. To help the engineers design best-in-class tires, computer simulation tools are used to predict the performance of the tires already during the early phases of the development process. These tools continuously evolve as computers become more performant and the simulation methods more refined. A very complex topic in this field is the simulation of snow. It is especially challenging due to the fact that snow material behavior can be very different depending on the meteorological conditions. This can for example be observed on a snowy road or a ski slope, where freshly fallen snow is often loose and soft while old snow is dense and hard. Depending on which state it is in, it will feel completely different to the person driving or skiing on it.

To be able to include snow in the computer simulations, mathematical models are required which describe the physical nature of snow. Depending on how detailed these models are the more difficult and time-consuming the simulations will become. A complex system like a tire rolling on snow includes both phenomena which appear on the microscale, like the interaction between the tire surface and the snow grains, as well as phenomena at the macroscale, like the deformation of the snow in the tire track. The goal of this project is to develop a snow model, which can simulate efficiently such a system on the macroscale while at same time incorporating to a large degree the processes occurring at the microscale. To achieve this, a so-called multiscale model will be used that was original developed for granular geomaterials. The direct benefit for Goodyear from this project is that it gets: (1) an improved model for snow and (2) the corresponding processes to calibrate the model against physical experiments. The former will allow Goodyear to better represent snow in their computer simulations while the later will make sure that the snow in the simulation behaves as the one found on a snowy road in winter. In addition to the actual academic relevance of such a challenging investigation, the potential applications of this extension go well beyond the tire-snow interaction, by encompassing for example natural hazards related to snow avalanches or permafrost melting driven rockfalls.

PhD – Domain Biology, Life Sciences & Medicine

Abstract

Microalgae are microscopic unicellular algae found in freshwater and marine environments. Microalgae serve a crucial role in our ecosystem as they perform photosynthesis, i.e., they are responsible for producing half amount of atmospheric oxygen, and at the same time, they utilise sunlight energy and carbon dioxide for their growth. Microalgae is a rapidly emerging biomass for the food, nutraceutical and cosmetic industry. That is mainly ascribed to their peculiar nutritional value. They are rich in proteins, essential lipids, dietary fibres (soluble and insoluble), and a plethora of bioactive molecules (e.g. carotenoids, chlorophylls and phycobiliproteins), which well document their active role in supporting human bodily functions.

Nowadays, several microalgae-based products are found in the market, such as food supplements (ground microalgae biomass in powder or tablet form) and cosmetics enriched with natural microalgal essential oils and pigments. Spirulina and Chlorella are the leading types of microalgae found in the market. Despite its crucial dietary role, the consumption of raw microalgal biomass may be nutritionally compromised. Many of the nutritionally appealing constituents are enclosed in very rigid cell wall structures leading to poor assimilation during the digestive process.

The term “probiotics” is used to denote living microorganisms (bacteria or yeasts) that confer beneficial effects when administered to the human host. Numerous health benefits such as regulation of the gut balance and immune system response mechanisms, relief of symptoms associated with food intolerance (e.g. lactose) or allergenicity, counteracting Irritable Bowel Syndrome (IBS) and travellers’ diarrhoea as well as preventive action against several forms of cancer have been associated with probiotics. Traditionally, probiotics are intaken via fermented foods such as yoghurt, kefir, sauerkraut, pickles, kimchi, kombucha etc. Modern lifestyle has given rise to novel concepts of intaking sufficient doses of living probiotics via specifically designed food supplements, including infant food formulas. Nevertheless, probiotics’ viability may be drastically reduced due to the countereffect of manufacture, long-termed storage and digestion (e.g. exposure to the stomach and intestine digestive fluids).

The cornerstone of the ALGPRO project is to design and study alternative innovation routes in the domain of personalised nutrition via joint research and development (R&D) actions between LIST and PM-International by adopting the current European Commission’s research and innovation policy plans for the sustainable transformation of the conventional food ecosystems (Food 2030, EU Green Deal, Farm-to-Fork).

ALGRPO project aims to provide a proof-of-principle insight into the ability of microalgal protein isolates to promote the biological activity of probiotic living cells (Lactobacilli) conveyed in the novel bespoke food supplement formulations. The project foresees the elucidation of the mechanisms involved in the action of microalgal proteins as regards to their: a) ability to preserving the viability of the probiotic cells under stress conditions (e.g. heat or osmotic shock, exposure to oxidative or low pH environments) typically encountered during manufacture, storage and household usage of food supplements, and b) ability to maintain their biological activity during the digestive process allowing a sufficient number of living probiotic cells to reach, adhere and become natural microbial colonisers of the colon microbial ecosystem.

It is expected that the ALGPRO project will create a dynamic field of research allowing the development of new generation nutrition products meeting the dietary needs of consumers adopting the well-being promoting, eco-conscious, clean label and society inclusive lifestyle (vegan, vegetarian, kosher etc.) concept.