File Name : fig.s1.jpg

Caption : fig. s1. the tc distribution statistics from the supercon database.5

File Name : fig.s2.jpg

Caption : fig. s2. predicted tc distribution and some of the best candidates in the materials project database. 6

File Name : fig.s3.jpg

Caption : fig. s3. neural network training, testing and setup. a the convergence of a model training. b performance of the dnn trained by afs-2 on a random test, the score of r2, rmse and mae on the test set are used as the performance indicator of the model. c deep neural network layer setting.

File Name : fig.s4.jpg

Caption : fig. s4. residual of dnn trained by afs-1(s1), the absolute error is mostly within 5k and it is distributed within 20k.

File Name : fig.s5.jpg

Caption : fig. s5. residual of dnn trained by afs-2(s2), the absolute error is mostly within 10k, and it is distributed within 25k.

File Name : fig.s6.jpg

Caption : fig. s6. in the interval of 0.44  0.03, each threshold sliding step is set to 0.004. by observing the confusion matrix, we find that by adjusting the threshold, the frequency of serious errors can be reduced while maintaining other scores, thereby reducing the model's serious errors in the classification task. 

File Name : fig.s7.jpg

Caption : fig. s7. according to the order of abcd, the manifold learning methods of principal component analysis (pac), multidimensional scaling (mds), t-distributed stochastic neighbor embedding (t-sne) and isometric feature mapping (isomap) show the dimensionality reduction visualization of the previous layer of the output layer of the deep neural network.7–9

File Name : fig.s8.jpg

Caption : fig. s8. the virtual sample prediction results with a hg, b pb, c ca, d ba, and e cu as independent variables by dnn (trained by afs-1), the blue and orange curves represent the highest and lowest predicted values, respectively.

File Name : fig.s9.jpg

Caption : fig. s9. the virtual sample prediction the highest tc with ba by rf model (trained by afs-1), it also found a dip in the 0.2~0.25 interval of ba weight, there may be a mysterious physical effect or possible wrong data in the data set.

File Name : fig.s10.jpg

Caption : fig. s10. the importance ranking of the afs-1 descriptors extracted from the rf model. the richness of the elements is shown in the high-resolution picture “elem.jpg”.