Equivariant Message-Passing Neural Networks and Probabilistic Diffusion Models for Materials Science
- Doctorant:
- Astrid Klipfel
- Co-directeurs de thèse :
- Yaël Frégier (LML)
- Adlane Sayede (UCCS)
- Co-encadrant de thèse :
- Zied Bouraoui
- Financement : ANR, Artois
- Thèse soutenue le :
- 22 déc. 2023
Hydrogen has the potential to be a cleaner alternative to other energy sources that produce high levels of greenhouse gases. However, to effectively reduce greenhouse gas emissions, we need to establish infrastructures to enable hydrogen production with lower emissions. One potential solution could be solar hydrogen production, which results in hydrogen with significantly lower greenhouse gas emissions. However, this process requires the development of new semiconductor materials with photocatalytic properties. Within the broad aim of discovering new semiconductor materials that can be used for photocatalytic water-splitting, various strategies mainly based on simulation or machine learning can be explored. While the vast majority of these methods have proven to be effective in organic chemistry, especially those based on message-passing neural networks, it remains unclear to what extent or in what way message-passing neural networks can be adapted to producing crystalline materials such as semiconductors. This thesis first proposes a physically inspired equivariant architecture of message-passing neural network architecture for crystal materials based on vector fields. These MPNNs are then used in generative models based on a probabilistic diffusion process, which is found to produce interesting crystalline materials. These generated models can then be helpful in high-throughput screening pipelines by selecting new candidate materials with the required characteristics for solar hydrogen production.