Fig. 2: Pre-training results.
From: A data-efficient foundation model for porous materials based on expert-guided supervised learning
a Pearson correlation coefficient of geometric features. b Coefficient of Determination (R2) score of geometric features. SpbNet is trained to predict different geometric features: (1) Accessible Surface Area (ASA) of different probes, (2) Void Fraction (VF) of different probes, (3) Pore Limited Diameter (PLD), and (4) Largest Cavity Diameter (LCD). Both ASA and VF are calculated using probes of 0.5 Å and 2.0 Å. Thus there are totally 6 different geometric features. These results are calculated from 100 Metal Organic Frameworks (MOFs) randomly selected in the pre-training validation set. c t-distributed Stochastic Neighbor Embedding (t-SNE) plot of Potential Energy Surface (PES) patches' feature colored by the contained number of atoms. From the lower left to the upper right, the number of atoms gradually increases. d t-SNE plot of PES patches' feature colored by Signed Distance Function (SDF). From the top left to the bottom right, the SDF gradually increases. For both t-SNE plots, PES patches are from 100 MOFs randomly selected from the validation set of the pre-training process. Source data are provided as a Source Data file.
