Nos tutelles



Accueil > Coin des étudiants > Stages Prolongeable par une thèse. > Tous les stages.

Unravelling the structure of complex hydrocarbons from their spectroscopic response by neural networks and machine learning

publié le

Growing evidence from astronomical observations suggests that carbon matter is present in interstellar media in large amounts, possibly under highly ordered form such as fullerenes. However, structural assignment is difficult in general because spectroscopic bands are usually not as resolved as they are for this highly symmetric molecule, and one current challenge in astrochemistry is to relate the measured spectra to molecular conformations, and to notably determine how ordered or amorphous hydrocarbon nanoparticles found in Space can be.

To address this issue from a theoretical perspective, we have generated large numbers of configurations of hydrocarbon nanoparticles and obtained their vibrational spectrum. The question now amounts to connect these two types of properties, and for this we propose to follow a neural network approach in which the main input variables will be some structural order parameters that characterize the nanoparticle conformations, the network providing a prediction of some main spectroscopic features. The order parameters include global shape parameters (gyration radius, asphericity), more local properties (coordination, hybridization level, size of aromatic domains) and the composition of the nanoparticle itself or hydrogen/carbon content. As outputs, the neural network will predict the propensity of the the input structure to exhibit vibrational lines at specific regions and corresponding to C-H vibrations, out-of-plane modes, etc.

Once trained on a set of available data, the neural network will be challenged on a broader set of configurations and its predictive capability assessed.

Florent Calvo, 04 76 51 45 92
LiPhy, 140 rue de la Physique, 38402 Saint Martin d’Heres