|DATE||October 28 (Thu), 2021|
|TITLE||Benchmarking Extended ACBN0 Method for Optical Properties of Porous Materials and Predicting Synthesizability of Porous Materials using PU Learning|
|ABSTRACT||Recently, the newly developed extended(e)-ACBN0 method which self-consistently determines both the on-site (U) and intersite (V) Hubbard interactions showed the excellent performance for the band gaps of low dimensional as well as bulk systems [Phys. Rev. Research 2, 043410 (2020)]. Here we present first-principles calculations of the optical band gaps of porous materials such as metal?organic framework (MOF), covalent organic framework (COF), and Zeolite using e-ACBN0 calculations compared to PBE, HSE, and GW calculations. For e-ACBN0 and GW calculations, we employ the Bethe-Salpeter equation (BSE) to accurately predict exciton binding energies of porous materials. According to our calculations, e-ACBN0+BSE generally underestimates the optical band gaps except a zeolite system (HEU). On the contrary, GW+BSE generally overestimates the optical band gaps of all porous materials with the mean absolute error (MAE) of 0.6 eV which is 0.25 eV smaller than that of e-ACBN0+BSE (0.85 eV). Interestingly, the exciton binding energies predicted by e-ACBN0+BSE and GW+BSE calculations are much larger than those of inorganic systems since the valence band maximum (VBM) and conduction band minimum (CBM) states of porous materials are spatially localized on the same subunit of the structure. For instance, the exciton binding energy of MOF5 obtained from GW+BSE reaches to 2.63 eV which is about four times larger than that of MoS2. We further find that HSE06 shows the best performance even though HSE does not account for electron-hole interactions indicating that the resultant cancellation of the errors very plausibly leads to fortuitous agreement with experiment in this case.
In part II, we introduce our machine learning model to predict the synthesizability of porous materials, which is determined by complex factors such as kinetics, temperature, and atmospheric conditions as well as the thermodynamic stability. Therefore, the synthesizability only based on the formation energy obtained from DFT calculations is often unpredictable. Considering this, we examine the synthesizability using the well defined porous materials database and machine learning classification model. In particular, we employ the positive and unlabeled (PU) learning to classify each porous material as synthesizable or unsynthesizable. For a structural description, we encode the crystal structures of porous materials using the crystal graph convolutional neural networks (CGCNN). According to our preliminary results, about 70 % of the synthesized porous materials are predicted as synthesizable. We expect that to optimize input parameters for PU learning will further improve the accuracy of our classification model.