Computational Modelling — Machine Learning-Based Prediction of Proton Conductivity in Metal-Organic Frameworks

4 measurement groups · 18 results

Reported values remain attached to the sample, method, conditions, extraction quality and source location that produced them.

Descriptor-based ML: ANN, Gaussian Process Regression, and XGBoost

curated 248-CIF proton-conductive MOF model/data set · Model

Input features include MOF RAC/geometric descriptors, RDKit guest descriptors, temperature, and relative humidity; train/test split by MOF structure with 5-fold cross-validation.

Context: model_system
Measurement source: 11283 · Machine Learning Model Construction · Figure 3a; Table 1

Property	Reported value	Normalised value	Uncertainty	Origin and quality	Source
descriptor + ANN test MAE	1.22 +/- 0.07	—	+/- 0.07	Table Exact Reported	11283 · Performance Comparison · Table 1
descriptor + GPR test MAE	1.2 +/- 0.04	—	+/- 0.04	Table Exact Reported	11283 · Performance Comparison · Table 1
guest descriptor count	199 features	—	—	Text Exact Reported	11283 · Machine Learning Model Construction
randomly divided XGBoost test MAE	0.28	—	—	Caption Exact Reported	S-7 · Supporting Information · Figure S3
randomly divided XGBoost training MAE	0.08	—	—	Caption Exact Reported	S-7 · Supporting Information · Figure S3
randomly divided XGBoost validation MAE	0.28	—	—	Caption Exact Reported	S-7 · Supporting Information · Figure S3
total feature count for descriptor model	375 features	—	—	Text Exact Reported	11283 · Machine Learning Model Construction
descriptor + XGBoost test MAE	0.98 +/- 0.07	—	+/- 0.07	Table Exact Reported	11283 · Performance Comparison · Table 1

XGBoost feature importance, PCA of transfer-learning representations, and attention-score analysis

curated 248-CIF proton-conductive MOF model/data set · Model

Feature importance across folds; PCA coloured by proton conductivity, RH, and temperature; attention examples for IKOTUZ and KUXREC.

Context: model_system
Measurement source: 11284 · Performance Comparison · Figures 5, S6, S7

Property	Reported value	Normalised value	Uncertainty	Origin and quality	Source
MOFs without open metal sites	approximately 175 No OMS	—	—	Visual Estimate Approximate	11285 · Performance Comparison · Figure 5g
MOFs with open metal sites	approximately 70 OMS	—	—	Visual Estimate Approximate	11285 · Performance Comparison · Figure 5g
XGBoost T+RH+Guest descriptor-combination MAE	MAE: 1.00 +/- 0.03	—	+/- 0.03	Figure Axis Rounded Reported	11285 · Performance Comparison · Figure 5b
XGBoost T+RH+Guest+MOF descriptor-combination MAEMarked as a best value within this paper	MAE: 0.98 +/- 0.07	—	+/- 0.07	Figure Axis Rounded Reported	11285 · Performance Comparison · Figure 5c
XGBoost T+RH descriptor-combination MAE	MAE: 1.09 +/- 0.04	—	+/- 0.04	Figure Axis Rounded Reported	11285 · Performance Comparison · Figure 5a

Transfer-learning feature-combination comparison

curated 248-CIF proton-conductive MOF model/data set · Model

Concatenation, Arrhenius-equation feature stacking, and element-wise addition were compared for the freeze transfer-learning model.

Context: model_system
Measurement source: 11284 · Performance Comparison · Figure 4

Property	Reported value	Normalised value	Uncertainty	Origin and quality	Source
transfer freeze Arrhenius-equation test MAE	approximately 1.16	—	—	Visual Estimate Approximate	11284 · Performance Comparison · Figure 4
transfer freeze concatenation test MAE	approximately 1.06	—	—	Visual Estimate Approximate	11284 · Performance Comparison · Figure 4
transfer freeze element-wise-addition test MAEMarked as a best value within this paper	approximately 0.91	—	—	Visual Estimate Approximate	11284 · Performance Comparison · Figure 4

Transformer transfer learning using MOF Transformer and ChemBERTa

curated 248-CIF proton-conductive MOF model/data set · Model

MOF CIF embeddings and guest SMILES embeddings combined with temperature and RH embeddings; freeze and unfreeze/fine-tune variants evaluated.

Context: model_system
Measurement source: 11283 · Machine Learning Model Construction · Figure 3b; Table 1

Property	Reported value	Normalised value	Uncertainty	Origin and quality	Source
transfer learning freeze test MAEMarked as a best value within this paper	0.91 +/- 0.04	—	+/- 0.04	Table Exact Reported	11283 · Performance Comparison · Table 1
transfer learning unfreeze fine-tuning test MAE	0.98 +/- 0.05	—	+/- 0.05	Table Exact Reported	11283 · Performance Comparison · Table 1