Design of magnetic MOF by machine learning methods – MagDesign

Two- and three-dimensional coordination polymers are conceptually the result of a fantastic Lego game. Indeed, their crystalline framework is the result of the arrangement of elementary components: a metal site, associated with a redox ligand, and a counter-ion inserted into the cavities of the metal-ligand framework. Exploring the possibilities of such an ensemble by usual methods, whether experimental or theoretical, is dizzying. A change of paradigm is thus required. In this project, we propose to develop a non-heuristic methodology, using machine learning methods, for the design and prediction of magnetic MOFs from a set of elementary components.

To reach this goal, we will develop an evolutionary algorithm dedicated to MOFs and interface it with USPEX, in order to predict the composition, structure and properties of magnetic MOFs. This algorithm will use a set of descriptors, chemical ones (to be determined) to quantify the metal-ligand bond strengths and the metallicity of the systems, and magnetic interactions. As the calculation of the latters is numerically expensive, we will replace it by an evaluation via a neural network. The determination of the best type of deep learning method to be used, of the structural or electronic descriptors of magnetism, the constitution of the training data set, will constitute the second part of this project. Finally, theoretical results will be tested upon chemical reality and the predictions compared to experiments. The synthesis, structural determination and measurement of the magnetic properties of first known, then predicted compounds will constitute the last part of this project.