All of the models trained during the research are available in this directory. Each model's name starts with either ecfp, for the linear models based on extended-connectivity fingerprints (ECFPs), or gnn, for the graph neural network (GNN) models. The name also indicates which subset of the Qin data the models were trained/tested on: either all of the data, or just the nonionic surfactants.
The sensitivity models' names also contain the syntax <num_splits>-splits-trial-<trial_idx>. The four benchmarking models are stored in ecfp-all, ecfp-nonionics, gnn-search-all and gnn-search-nonionics. The GNN benchmarking folders also contain the best hyperparameters found during the Hyperband search (best_hps.json).
All of the GNN models contain logs subdirectories with data that can be read by Tensorboard:
tensorboard --logdir logs/Each subdirectory contains the final performance metrics and the predictions for the model at the end of training. The names are generally self-explanatory, but note that the uq_*.csv files in the gnn subdirectories refer to the results from the models with uncertainty quantification.