Ammonia is a widely used chemical, mainly as a fertilizer in agriculture. Its production currently relies on an industrial process, requiring high temperature and pressure conditions and thereby emitting a non-negligible quantity of CO2. It is however known that the reaction can happen at normal temperature and pressure conditions in presence of proper catalysts.
This project focuses on leveraging Machine Learning (ML) techniques to inverse-design materials with the desired catalytic properties.
Conceptually, a material can be seen as a repetition of a fixed box (containing atoms) along the x-, y-, z-axes. Depending on its composition (the atoms’ types), its configuration (the atoms’ positions) and its shape (e.g. cubic or parallelepipedic), the resulting material can exhibit very different properties.
The search-space over all possible boxes is both immensely-vast and discrete, making direct optimization difficult and random navigation extremely expensive. This project is about developing data-driven techniques that allow for generation of promising boxes, that in turn lead to materials with interesting properties. Practically speaking, this means that given a property of interest, we would like to be able to generate boxes that are likely to have those properties.
More specifically, we aim at solving this inverse problem by employing deep generative models, which have the ability to represent distributions of high-dimensional objects (i.e. materials). By conditioning those distributions on properties of interest, this would allow us to generate compounds with desired properties.
Picture: adapted from [Sanchez-Lengeling, B., & Aspuru-Guzik, A. (2018). Inverse molecular design using machine learning: Generative models for matter engineering. Science, 361(6400), 360-365.]