Fig. 3: Prediction performance of DeepSorption on CoREMOF and hMOF datasets.

a, b The correlations between true adsorption uptake and predicted adsorption uptake on test sets implemented by DeepSorption network on CoREMOF-CO₂ (a), hMOF-CO₂ (b) tasks. c The distribution of absolute errors on test sets between true adsorption uptake and predicted adsorption uptake on CoREMOF-CO₂ based on different models, including DeepSorption (Matformer+KCL), RAC_MLP (Multilayer perceptron based on RACs descriptors), CGCNN (Crystal Graph Convolutional Neural Network⁴⁷), MOFNet¹⁹, MBTR_MLP (Multilayer perceptron based on MBTR descriptors), GEO_MLP (Multilayer perceptron based on geometric structure descriptors). d The coefficient of determination R² on test sets of co-learning knowledge, including LCD (largest cavity diameter), PLD (pore limiting diameter), D (density), ASA (accessible surface area), VF (void fraction), AV (accessible volume) on CoREMOF-CO₂ using DeepSorption, CGCNN, MBTR_MLP, RAC_MLP and LSTM (Long Short-Term Memory⁴⁸) models. e, f The coefficient of determination R² and mean absolute errors (MAE) of different models on CoREMOF-CO₂ (e) and hMOF-CO₂ (f) tasks on test sets.