ExperimentalDatasetConstructionResults_BuildingInformativeMaterialsDatsetsBeyondTargetedObjectives

The CSV files in this repository contain the results of all active-learning runs performed using the sysTEm dataset as the candidate pool. These runs correspond to the dataset construction strategies and random initializations studied in the paper “Building Informative Materials Datasets Beyond Targeted Objectives.” The folder names indicate the dataset construction strategy used: grow_random: random sampling grow_QBC: query-by-committee (QBC) with one target grow_2QBC: QBC with two targets grow_3QBC: QBC with three targets grow_EA_2Obj_TwoTargets_QBC_ONLY: QBC with two targets using NSGA-II, without feature diversity grow_EA_3Obj_ThreeTargets_QBC_ONLY: QBC with three targets using NSGA-II, without feature diversity grow_EA_DeltaDiversity: QBC with one target plus feature diversity using NSGA-II grow_EA_3Obj_TwoTargets: QBC with two targets plus feature diversity using NSGA-II grow_EA_4Obj_ThreeTargets: QBC with three targets plus feature diversity using NSGA-II grow_multiobj_all: QBC with four targets Naming conventions used in the file names: el: electrical conductivity seeb: Seebeck coefficient ThC: total thermal conductivity zT: thermoelectric figure of merit xgb: XGBoost performance results rf: Random Forest performance results These files report the model performance obtained across acquisition iterations for each strategy, target combination, model type, and random initialization.