Thanks Daniel. This is awesome and very much appreciated. I think the performance of the predict functions might just roughly follow the distribution of the data (i.e. majority of the data has EF ~ 60% +- 5%). But, the method from your SYNASC paper is definitely really great. If interested, analysis of the top NDSB2 solutions (all used CNNs in workflow) here. Thanks again!

Do you have a repository with the 30x32x32 scans? Not necessarily all 11000 of them, but maybe 1000 or so? (If all are available somewhere that would be fine too.)