Fig. 5: Label Efficiency.
From: A data-efficient foundation model for porous materials based on expert-guided supervised learning
a Performance of models under different volume of fine-tuning data for H2 Uptake. b Performance of models under different volume of fine-tuning data for H2 Diffusion. Models are fine-tuned on 5 K, 10 K, 15 K, and 20 K data, which are randomly selected from the whole training set. Figures demonstrate the Mean Absolute Error (MAE) of different models under different fine-tuning data volume. SpbNet is compared to MOFTransformer20 (mt) and PMTransformer21 (pm). SpbNet is pre-trained on 100 thousand MOFs. MOFTransformer is pre-trained on on MOF datasets of varying size (100 thousand, 500 thousand, 1 million). PMTransformer is pre-trained on 1.9 million porous materials. Part of results are from the figure of MOFTransformer20. Source data are provided as a Source Data file.
