• Co-Advisor :
  • Yaël Frégier (LML)
  • Adlane Sayede (UCCS)
  • Funding : ANR, Artois

Ongoing advancement in progress in artificial intelligence (AI) and in particular structured and data-driven machine learning (ML) begins to motivate great interests in material science. For example, different material properties of stoichiometric inorganic crystalline compounds were predicted using ML models. Moreover, AI shows great power in helping materials design and synthesis. In the present project, we will develop AI models for predicting electronic properties (band-gap) of semiconducting electrodes for photoelectrochemical energy conversion. In particular, we challenge the problem of investigating novel chemical compositions with stable crystals and desired band-gaps. Traditionally, density functional theory (DFT) plays a central role in the prediction of chemically relevant compositions with stable crystals and desired band-gap. However, the DFT calculations are computationally expensive, and it is not adequate to apply it to test all possible randomly generated structures. A number of machine learning approaches were suggested to facilitate the search for novel stable compositions, among them, one can cite Generative Adversarial Networks (GANs). Thanks to the available large amount of high-quality computational materials-science public data (more than 50 000 000 instances), structure-property (band-gap, for instance) relationships can be used to train a generic model that could then be used to train a generator specifically to semiconducting electrodes for photoelectrochemical energy conversion. In our approach, crystal structures will be molded by means of graph theory instead of using the usual lattices and space groups to describe crystals. This approach uses quotient graphs and nets, where atoms are mapped onto nodes and bonds on to edges. These nets are then described by so-called quotient graphs that have the advantage of being finite and can, therefore, be manipulated more easily. To this end, we will investigate some variant of Graph Convolutional Networks (GCNs) to learn a vector representation of the nodes that occur in graph structure which is suitable for predicting new structure. GCNs are a generalization of Convolutional Neural Networks (CNNs). GCNs can learn to extract features from the given node representations and compose these features to construct highly expressive node vectors.