(611i) Role of Pore Chemistry and Topology in the CO2 Capture Capabilities of MOFs: Molecular Simulation and Machine Learning | AIChE

(611i) Role of Pore Chemistry and Topology in the CO2 Capture Capabilities of MOFs: Molecular Simulation and Machine Learning


Gomez Gualdron, D. - Presenter, Colorado School of Mines
Anderson, R., Colorado School of Mines
Rodgers, J., Colorado School of Mines
Open-framework materials (OFMs) such as metal-organic frameworks (MOFs) are promising for applications such as CO2 capture. The challenge is that with millions of possible “crystal topology + composition” combinations it is very difficult to identify the best possible MOFs, to clearly elucidate MOF performance limits, and to isolate the impact of different MOF features on performance solely doing experiments. High throughput computational screening (HTCS) have emerged as a powerful tool to study large numbers of MOFs. However, despite vast amounts of CO2 adsorption data collected in various HTCS efforts, the role of the MOF pore chemistry and topology remains to be clearly elucidated due to factors in earlier HTCS studies such as i) limited number of topologies considered, ii) limited number of functional groups considered, iii) incomplete MOF series, and iv) accuracy uncertainties for the description of electrostatic interactions.

In this work, we leverage our tools to construct MOFs using the TobaCCo code[1] and to inexpensively calculate charges that describe well electrostatic interactions using MBBB charges[2] to study the adsorption of CO2 and CO2/N2 and CO2/H2 mixtures in a “complete” population of MOFs. The “complete” population of MOFs contains all possible “crystal topology + functionalization” combinations for 15 topologies and 13 functionalized building blocks, which has allowed us to comprehensively examine the role of pore topology and chemistry on the CO2 capture capabilities of these open framework materials. Then, we use this knowledge to provide “intuitively derived” descriptors based on data from DFT calculations and GCMC simulations to train machine learning algorithms to quantitatively predict CO2 capture properties of the studied materials without the need of adsorption simulations.

We evaluated the performance of six different machine learning algorithms, including artificial neural networks (ANNs) and gradient boosting machines (GBMs). We found that, for similar computational cost, gradient boosting machines consistently outperformed neural networks in prediction accuracy. Finally, by using individual conditional expectation (ICE) plots we were able to further analyze the response of CO2 capture “metrics” to changes in MOF features, including chemistry and topology, allowing us to predict the optimal values of these features for future optimization of MOF design.


  1. Cryst. Growth. Des. 2017, 17 (11), pp 5801-5810
  2. J. Chem. Theory Comput., 2018, 14 (1), pp 365–376