Navigating through digital reticular chemistry

Reticular chemistry provides us with a practically continuous design space along multiple dimensions: We can tune the chemistry on the atomic scale, tune the sizes and connectivity of pores, or engineer patterns in the repetition of building blocks.
To find our way in this MOF landscape, it would be nice to have a map that guides us. In this presentation, we show that machine learning can provide us with such a map [1]. We can use machine learning to learn patterns that are tacit in a large number of dimensions of this chemical space and then use it to guide the reticular design.
The simplest application of this navigation system is to predict properties that are hard to predict with conventional quantum chemistry or molecular simulation alone [2, 3] — including in which net MOFs are most likely to self-assemble.
Once we have this in place, we can use it to most efficiently gather information about structure-property-function relationships. A key difficulty here is, however, that we often have to deal with multiple, often competing objectives. Interestingly, one can show that using a geometric construction one can also effectively, and without bias, use machine learning to dramatically accelerate materials design and discovery in such a multiobjective design space [4].
It is important to realize, however, that machine learning relies on data that a machine can use [5]. Towards this goal, we need to develop infrastructure to allow for the capture without overhead while providing chemists with tools that simplify their daily work [4, 5].
However, it is also interesting to realize that also machine learning itself can be used to curate data [2]. For instance, we could show that by simply analyzing the training dynamics one can identify chemically invalid structures in widely used datasets.
All these findings provide us with a toolbox that can change how we design reticular materials. We can capture all the relevant data using novel tools, use them to inform our machine learning systems, which then inform us which coordinates of reticular space to explore. This might allow us to tackle larger and more complex problems at a higher pace, and hopefully, learn more about the patterns underlying the reticular design space.