Extended Data Fig. 2: Transfer learning of ion association in solution.
From: Machine-guided path sampling to discover mechanisms of molecular self-organization
Committor cross-correlation plots for models obtained by transfer training on different ionic species with varying number of training shooting points. The original model was trained on LiCl and is the same in all cases (see Fig. 1 for a committor cross-correlation plot of the original model predicting for LiCl). The model is of architecture ‘ResNet I’ and uses the inputs ‘SF longranged II’ (see Supplementary Tables 5 and 6). Transfer training was done by randomizing the last layer of the original model with a single neuron and then training on the specified number of shooting points with data for each ionic system separately. The average of the sampled committors (blue line) +/- one SD (orange shaded) were calculated by binning the sampled committor in the range of the learned committor indicated by the vertical steps. For reference, the red line indicates the identity.
