1 of 1

An ecosystem for digital reticular chemistry

Scientific Achievement

Digital chemistry has shown great progress in energy-related applications ranging from carbon capture to gas separation. While machine learning (ML) holds promise to accelerate this field, there is still a lack of good ML practices and datasets in this area. We present the fundamentals of an ecosystem of datasets, methods, and good practices to advance this ecosystem.

Significance and Impact

The ecosystem we have created provides increased accessibility to machine learning for reticular chemistry and beyond without compromising on rigor, especially for less experienced users. This will allow for a closer coupling of data-driven materials design and the synthesis and characterization of (in silico generated) materials.

Technical Approach

We show key pitfalls, such as researchers having significant data leakage between training and test sets, preventing direct comparisons of modeling approaches.
We identify that models are often not validated correctly (and through different protocols), and develop better validation criteria for ML models in this field.
We provide new ML methods that can be used directly on the newly curated datasets.

K. Jablonka, A. S. Rosen, A. S. Krishnapriyan, B. Smit, ACS Central Science (2023).

Better validation metrics for splitting datasets

TALKING POINTS:

The field of reticular chemistry aims to develop materials with improved gas adsorption, better electronic properties, heat capacities, and stability.
Reticular chemistry, the science of constructing extended crystalline structures from molecular building blocks, gives scientists a unique playground for material design and discovery as it gives access to a practically infinite-dimensional design space across many length scales: One can architect the pore, functionalize the building blocks, or even encode chemical sequences across unit cells.
This work is motivated by the observation that the full potential of data science in this field has not yet been achieved. Even though all works operate on the same class of materials and often use related machine-learning approaches, these works are hardly comparable or replicable and only implementable by experienced groups.
We introduce a package that provides increased accessibility to machine learning for reticular chemistry and beyond without compromising on rigor, especially for less experienced users. This will allow for a closer coupling of data-driven materials design and the synthesis and characterization of (in silico generated) materials.

METADATA:

Name of the associated awarded project: SciDac-5 Institutes RAPIDS2

PI name(s):

Name of the program manager:

CITATIONS:

K. Jablonka, A. S. Rosen, A. S. Krishnapriyan, B. Smit, ACS Central Science (2023).