Numerous fields require the solution of increasingly large-scale optimization problems, which involve large amounts of high-dimensional data and large complex models, notably deep neural networks. The trend to enhance the performance of such models has been to progressively increase their complexity and to feed them an increasingly larger amount of data, so that their training and deployment requires extensive computational resources. It is thus of crucial importance to make deep learning more parsimonious, by reducing its data-hungry nature and its dependence on large DNNs. The aim of this project is to develop optimization methods that are computationally efficient, parsimonious of resources and mathematically sound. The key ingredient to achieve this goal will be to develop methods able to build, exploit or discover hierarchical structures in the problem at hand. These structures allow for representations of a problem at different scales of size, accuracy and/or resolution, each of them corresponding to lower and lower costs/amount of resources, thus allowing multiple flexible representations that can achieve a flexible compromise between accuracy and parsimony. A clever use of this variety of models will allow to achieve the efficiency and parsimony targets without compromising accuracy.

The prospect of mastering and exploiting the motion of molecules is a fundamental challenge related to the development of molecular architectures featuring tunable properties. In this context, the ElectroMagnet project aims at developing metal-organic molecular systems for which large amplitude reorganizations can be used to control the spin state of magnetically active metallic centers. According to this strategy, the magnetic state of the metal will be defined by its coordinated ligands whose number can be controlled by electron transfer centered on key organic fragments. The proposed concept relies in particular on the pi-dimerization of viologen cation radicals as the driving force of structural reorganizations.

One great challenge nanoscience is facing is the difficulty to transpose molecular-scale phenomena into macroscopic properties finding application in everyday-life devices. A way to address this issue is to develop metamorphic molecular systems for which an external stimulus triggers a drastic structural reorganization. By controlling supramolecular self-assembly of metamorphic building blocks it is indeed possible to develop responsive materials for which properties at the macroscopic level can be modulated. Based on this strategy ChiroSwitch will bridge the gap between chiroptical properties observed in solution for Circularly Polarized Luminescence (CPL) switches and their use in device-like systems. Despite being crucial steps toward application in photonics and optoelectronics, modulation of CPL properties has rarely been achieved on-surface and switches responding to an electric stimulation remain almost unexplored. The major breakthrough of ChiroSwitch relies on the hypothesis that metamorphic processes occurring in supramolecular assemblies can be used to achieve on-surface modulation of chiroptical properties. To succeed in this ambitious goal, the main objective of this project is to design responsive chiroptical switches whose supramolecular self-assembly and CPL properties on surfaces can be controlled with optical or electrical stimulations. ChiroSwitch will be constructed around three axes of increasing difficulty and risk: 1) The metamorphic processes in self-assembled monolayers of bis-viologen hinges will be investigated on-surface by Scanning Tunnelling Microscopy. 2) CPL switches constituted of bis-viologen hinges and boron-based emitters featuring two intense CPL states will be studied in solution. 3) On-surface CPL modulation will be achieved via metamorphic processes occurring in the self-assembled monolayers of CPL switches. Knowledge acquired thanks to ChiroSwitch will serve the development of CPL devices with promising application for quantum cryptography, memory and computing apt to address the exponential growth of numerical data to be safely stored and processed.

The timing of the emergence and the early evolutionary stages of life on Earth remains enigmatic. Organic matter preserved in Archean (4.0-2.5 Ga) sedimentary rocks may record these early evolutionary stages, but organic traces of life in such ancient rocks are challenging to distinguish from non-biological (abiotic) organic structures that may display similar morphological and geochemical features. An important source of abiotic organic matter on the Archean Earth, not yet investigated, could come from atmospheric synthesis. Yet, very little is known about the geochemical fate of such Archean abiotic aerosols during geological burial, hence, about their geochemical signatures in Archean rocks. Therefore, I propose to tackle this issue by combining laboratory experiments to synthesize analogues of Archean organic aerosols and to submit them to simulated geological burial processes. Analogues of Archean organic aerosols will be synthesized through cold plasma discharge experiments in gaseous mixtures thought to compose the Archean atmosphere (N2, CO2, CH4). The synthesized organic aerosols will then be submitted to simulated geological burial processes in sapphire anvil cells coupled with UV-resonant Raman spectromicroscopy to investigate the morphological and molecular transformations of both organic and mineral phases during the experiments. The morphology and the elemental and molecular compositions of the experimental products will be further characterized using SEM, TEM, and XANES microspectroscopy. Such experimental constraints on the morphological, elemental and molecular signals that may have been left by abiotic atmospheric synthesis in the Archean geological record of organic matter are mandatory to further distinguish biotic from abiotic Archean organics, and mitigate controversies about traces of early life activity on Earth.