The aim of the project is to provide low-cost and high efficiency tandem cells grown on crystalline silicon (c-Si) substrates, with merging both the monocristalline Si approach with the high-efficiency monocristalline multijunction approach based on III-V materials. These CPV cells will be used under natural lighting and under low light concentrators (100 suns) developed by IRDEP-CNRS, and benchmarked under medium concentration by Heliotrop sas. The PV cells efficiency is one of the most important parameters for the final cost of electricity, since it impacts directly the ratio between produced energy and production cost. With 22% efficiency modules based on c-Si, the technology seems to reach its limits. To increase further the efficiency of c-Si cells and modules, going to multijunction devices (association of two different absorbing layers in the same cell) seems to be the obvious choice. While many projects tend to focus on all silicon technology, best high bandgap cells are yet based on III-V compounds. This project proposes to demonstrate the proof-of-concept for a monolithic integration of high efficiency multijunction CPV device on a low cost monocristalline silicon substrate upon which a III-V lattice-matched material will be grown using molecular beam Epitaxy (MBE). This Lattice-Matched heterostructure with its very low structural defect densities (Dislocations, AntiPhase Domains, point defects) will be capable of sustaining III-V high performing PV devices onto silicon with long life-time. This novel route overcomes the problems of high cost substrates (as compared to Ge or III-V substrates used currently for this kind of CPV), the killer structural defect formation and reliability issues of lattice mismatched systems (metamorphic approach) and the low reliability and low lifetime of hybrid techniques (such as wafer bonding). The integration of photovoltaic functions onto a single silicon substrate will also achieve a reduction in the use of III-V based semiconducting materials in high-efficiency multijunction CPVs. The two main scientific and technologic objectives of the project are : 1) The achievement of GaAsPN (1.7 eV) single cell on Si (with a 15% efficiency under low concentration, i.e. 100 sun). 2) The demonstration of a high efficiency and low cost multi-junction solar cell: GaAsPN pn cell at 1.7 eV on Si pn cell at 1.1 eV (25% efficiency under low concentration, i.e. 100 sun, as a first step towards very high-efficiencies >30%) Lattice-matched layers and slightly tilted substrates are used to overcome the two main difficulties faced by the growth of III-V materials on silicon substrates: misfit dislocations and antiphase lattice defects, in order to obtain defect-free III-V materials and to get large minority carrier diffusion lengths for the PV applications. The PV devices will consist in high efficiency tandem cells III-V/Si double-pn-junctions separated with a Buried Tunnel Junction. The final structure will include a first bottom Si pn (1.1eV low gap) grown on the Si substrate, then a thin GaP layers is grown by MBE to prevent structural defects formation, a top cell GaAsPN pn (1.7eV large gap) junction is then grown on top of it. The project relies on a high quality consortium which brings together six french partners, and an associated European partner, with high, established competence and complementary methodology and expertise in their fields and leading appropriate workpackages: FOTON (growth of III-V materials), INL (Si-based PV technology), CEMES-CNRS (structural characterizations), IRDEP-CNRS (research in PV development), EDF R&D (a European leader in the Energy sector), HELIOTROP (French manufacturer of high concentration photovoltaic modules (HCPV)) and AALTO (a Finnish associated academic partner specialised in point defects analysis). The partners are active in European research consortia and in networks of excellence and they drive many projects on the national and international level.

ROOT will develop numerical methods for solving important problems in computer graphics and vision via the optimal mass transport theory. These problems relate to histograms, and can be advantageously written and solved via regressions with loss functions involving optimal transport. These problems include data fitting, supervised learning or statistical inference. Histograms are frequently encountered in computer graphics (e.g., reflectance functions, color palettes, distance histograms etc.) and vision (e.g., SIFT or HoG descriptors). But in many cases, they are treated simply as Euclidean vectors to fit the existing learning machinery, or other ill-suited metrics are used to compare and manipulate them. Meanwhile, the increasingly popular optimal transport theory considers histograms as piles of sand which physical motion requires effort. Based on this principle, optimal transport offers a framework with a meaningful way of comparing histograms as the amount of work required to reshape a pile of sand into another. It also proposes a way of interpolating histograms as the intermediate histograms produced during this motion. The ROOT project explores the use of optimal transport, but, as a metric in the context of inverse problems for graphics and vision, and machine learning. However, this theory remains costly in practice, and practical optimal transport solvers will also be explored. Challenges addressed by ROOT include: - Computationally efficient optimal transport solvers between histograms. Since our regression tools will make repeated calls to optimal transport solvers during the optimization iterations, this step should be made computationally efficient. We will investigate methods based on Voronoi / Power diagrams, linear programming and entropy regularizations via Bregman projections. We will strive to offer tools for high-dimensional scattered data, as well as simpler histograms living on low-dimensional grids, considering both Eulerian as well as Lagrangian approaches. - Solving important inverse problems in graphics and vision with optimal transport. These problems will be cast as regressions within the optimal transport framework. They will be used for inference, supervised learning and fitting. Typical computer graphics applications include fitting reflectance data to analytical models, inferring missing values for captured or measured data (reflectance, 3d geometries), and learning color histogram manifolds for image or video color manipulation. Typical computer vision applications include the inference of parameters in reflectance models based on images, and object recognition given sparse databases ROOT will offer a public open-source library for efficient histogram regression.

Artificial Intelligence (AI) algorithms, and in particular Deep Neural Networks (DNNs), typically run in the cloud on clusters of CPUs and GPUs. To be able to run DNNs algorithms out of the cloud and onto distributed Internet-of-Things (IoT) devices, customized HardWare platforms for Artificial Intelligence (HW-AI) are required. However, similar to traditional computing hardware, HW-AI is subject to hardware faults, occurring due to manufacturing faults, component aging and environmental perturbations. Although HW-AI comes with some inherent fault resilience, faults can still occur after the training phase and can seriously affect DNN inference running on the HW-AI. As a result, DNN prediction failures can appear, seriously affecting the application execution. Furthermore, if the hardware is compromised, then any attempt to explain AI decisions risks to be inconclusive or misleading. One of the overlooked aspects in the state-of-the-art is the impact that hardware faults can have in the application execution and the decisions of HW-AI. This impact is of significant importance, especially when HW-AI is deployed in safety-critical and mission-critical applications, such as robotics, aerospace, smart healthcare, and autonomous driving. RE-TRUSTING is the first project to include the impact of HW-AI reliability on the safety, trust, and explainability of AI decisions. Typical reliability approaches, such as on-line testing and hardware redundancy, or even retraining, are less appropriate for HW-AI due to prohibited area and power overheads. Indeed, DNNs are large architectures with important memory requirements, coming along with an immense training set. RE-TRUSTING will address these limitations by exploiting the particularities of HW-AI architectures to develop low-cost and efficient reliability strategies. To achieve that, RE-TRUSTING will develop fault models and perform a failure analysis of HW-AI with the aim to study their vulnerability towards explaining the HW-AI. Explainable HW-AI signifies reassuring that the HW-AI is fault-free, thus neither compromising nor biasing the AI decision-making. In this regard, RE-TRUSTING aims at bringing confidence into the AI decision-making by explaining the hardware on which AI algorithms are being executed.

Tool use behaviors have evolved not only in primates but also in some birds (e.g. parrots and crows) and even in some teleost fish (e.g. wrasses and cichlids). As birds and teleosts do not have a six-layered neocortex, these similar cognitive functions must have evolved independently in the lineages of mammals, birds, and teleosts (convergent evolution). Our research project aims to identify the necessary conditions for the emergence of such cognitive abilities during vertebrate evolution. The ability for problem solving and sequential object manipulation is the prerequisite condition for evolving tool use ability. The EVONECTOME project investigates the connectivity (TASK 1) and functions (TASKs 2-4) of associative brain areas involved in a problem solving task and sequential object manipulation in parrots and cichlids. We propose to use Pyrrhura molinae as a parrot model and Amatitlania nigrofasciata as a cichlid model, which are optimal for combining anatomical and behavioral examinations. TASK 1: We will take a connectomic approach using ex-vivo diffusion MRI to visualize neuronal networks with regard to putative associative areas in the pallium. The main goal is to reveal the connectivity of the associative areas, notably with premotor/motor areas. The results will verify whether there is a shared network logic for cognitive-motor integration in primates, parrots, and cichlids. TASK 2: To examine problem solving and object manipulation abilities of parrots and cichlids, we will perform two different behavioral tasks: a puzzle box task and operant conditioning tasks that require object manipulation as responses. After the behavioral tests, the brain areas involved in these behaviors will be examined in TASK 3 and TASK 4. TASK 3: Neuronal activation during the behavioral tasks will be visualized by functional MRI, using Manganese-enhanced MRI (MEMRI). By detecting large-scale brain activity during the behavioral tasks (TASK 2), we expect to identify brain areas involved in problem solving and object manipulation. TASK 4: To assess the role of dopamine (DA) neurotransmission in the behavioral tests, we will destruct DA fibers in the associative areas by a local injection of 6-hydroxydopamine (6-OHDA). DA is known to play a critical role in various higher-order cognitive functions in mammals. If DAergic disruption in non-homologous brain areas leads to the same behavioral effects in mammals, birds, and teleosts, this will further support the importance of DA in the convergent evolution of higher-order cognitive functions. Altogether, our study will give an insight into how morphologically diversified nervous systems achieved similar cognitive functions during evolution. If the same network pattern emerged independently in primates, parrots, and cichlids, this would indicate the existence of a limited degree of biological freedom (high constraints) for the evolution of intelligence in vertebrate brains.