Speaker
Description
Imagine a computer dreaming up a porous material with mathematically optimized properties for a specific application. In this presentation, I will share some ideas on how to practically realize this concept.
The exponential growth of available and hypothesized porous materials, including Metal-Organic Frameworks (MOFs), has necessitated a fundamental shift in our approach to selecting porous materials for specific applications. Computational screening methods have emerged as a necessary step to identify the most promising candidates before committing to costly experimental efforts.
Machine learning (ML) methods can significantly accelerate computational screening protocols. A typical application of ML models involves training them to predict material properties from features such as surface area or more complex characteristics, thus partially reducing the need for time-consuming molecular simulations. Essentially, the ML model facilitates the process of sifting through millions of materials to identify promising ones.
A new emerging idea is to discard direct screening approaches and instead develop methods that enable a computer to "hallucinate" new structures, guiding them towards a pre-specified objective function. This approach is not tied to any specific material database and, we speculate, could be a much more efficient way to search for materials. Implementing this idea, however, requires overcoming several challenges associated with material representation, interpretability of the model, and the physical realism of generated structures. Nonetheless, it may also create the possibility of discovering entirely new materials with unprecedented properties and novel adsorption phenomena.
