Generative models are powerful methods that allow the synthesis of new data by approximating the probability distribution that describes a dataset. During the past decade, several works have demonstrated their utility to a wide range of applications, including in materials science for the design of new structures. In this thesis, we are interested in applying these techniques to the generation of porous media, constrained by different physical properties related to the characterization of lithium-ion batteries.
Our research is mainly focused on conditional models, combined with disentanglement approaches, with the objective of independently controlling the different properties of interest.
First, we propose a VAE-based method that improves weakly supervised disentanglement methods in a general context. Then, we couple disentanglement to conditional modelling and propose a GAN-based architecture that generates 2D porous media, able to disentangle scalar properties, namely the porosity and the contact surface. Next, we adapt our model to the generation 3D porous media, and propose the disentanglement of more complex and vectorial properties. Finally, we also implement a novel disentanglement metric, focused on the structure of materials, that is more adapted to our data.
