Table 1.

Summary of theoretical, QSPR, and machine learning (ML) models investigated in the literature

Database	Features	Model	$R^2$	Ref.
600	chemical groups	group contributions approach	N/Aa	18
32	an effective mobility value	single adjustable parameter	N/Ab	28
113	quantum chemical descriptors	artificial neural networks	0.955c	42
37	quantum chemical descriptors	support vector regression	0.97	43
251	Descriptors	computational neural networks	0.96	51
389	descriptors	support vector regression	0.78	52
133	descriptors	random forest	N/Ad	53
88	descriptors	multi-layer perceptron neural network	0.96	54
77	descriptors	support vector machine (SVM)	0.92	55
54	descriptors	artificial neural network	0.91	56
52	descriptors	artificial neural network	0.978e	57
451	hierarchy fingerprint	Gaussian process regression	0.94	38
751	hierarchy fingerprint	Gaussian process regression	0.87	37
1,321	hierarchy fingerprint	Gaussian process regression	0.92	39
5,917	combined fingerprint	Bayesian linear model	0.916f	41
331	SMILES-based binary images	convolutional neural network	N/Ag	49
234	SMILES-based binary images	fully connected neural networks	N/Ah	50
6,923 + 5,690 + 1 million	descriptors Morgan fingerprint SMILES-based binary images	lasso regression deep neural network convolutional neural network	0.80 0.85 0.87	this work

N/A, not applicable.

^aAbout 80% of the calculated T_g values differed less than 20 K from the experimental values.

^bOnly root-mean-square error of 13°C was reported for all 32 alkylated conjugated polymers.

^cR = 0.955 was reported for the prediction set.

^dOnly root-mean-square error of 4.76 K was reported for the test set of the model.

^eR = 0. 978 was reported for the test set.

^fR = 0. 916 was reported for the test set.

^gThe model performance was evaluated by relative error of 3%–8%.

^hThe model performance was evaluated by average relative errors of ∼3%.