The inert gases Xe
and Kr mainly exist in the used nuclear fuel
(UNF) with the Xe/Kr ratio of 20:80, which it is difficult to separate.
In this work, based on the G-MOFs database, high-throughput computational
screening for metal–organic frameworks (MOFs) with high Xe/Kr
adsorption selectivity was performed by combining grand canonical
Monte Carlo (GCMC) simulations and machine learning (ML) technique
for the first time. From the comparison of eight classical ML models,
it is found that the XGBoost model with seven structural descriptors
has superior accuracy in predicting the adsorption and separation
performance of MOFs to Xe/Kr. Compared with energetic or electronic
descriptors, structural descriptors are easier to obtain. Note that
the determination coefficients
R
2
of the
generalized model for the Xe adsorption and Xe/Kr selectivity are
very close to 1, at 0.951 and 0.973, respectively. In addition, 888
and 896 MOFs have been successfully predicted by the XGBoost model
among the top 1000 MOFs in adsorption capacity and selectivity by
GCMC simulation, respectively. According to the feature engineering
of the XGBoost model, it is shown that the density (ρ), porosity
(ϕ), pore volume (Vol), and pore limiting diameter (PLD) of
MOFs are the key features that affect the Xe/Kr adsorption property.
To test the generalization ability of the XGBoost model, we also tried
to screen MOF adsorbents on the CO
2
/CH
4
mixture,
it is found that the prediction performance of XGBoost is also much
better than that of the traditional machine learning models although
with the unbalanced data. Note that the dimension of features of MOFs
is low while the quantity of MOF samples in database is very large,
which is suitable for the prediction by model such as XGBoost to search
the global minimum of cost function rather than the model involving
feature creation. The present study represents the first report using
the XGBoost algorithm to discover the MOF adsorbates.