Dielectric Polymer Genome: Integrating Valence-Aware Polarizable Reactive Force Fields and Machine Learning

Interpretable models for extrapolation in scientific machine learning

“…Before tting the models, each dataset was scaled using the Scikit-learn RobustScaler method. Model performance was quantied using non-dimensional model error (NDME), 59 calculated as the ratio between root-mean-square error (RMSE) between predicted values and ground truth values in the test set, and the standard deviation in the ground-truth values of the test set. NDME was used as a performance metric to ensure that model performance could be compared across datasets and target properties with varying units and magnitudes, where NDME of 0 corresponds to a perfect model, and NDME of 1 corresponds to a model with average prediction error which is equal to standard deviation in the ground truth values.…”

Dielectric Polymer Property Prediction Using Recurrent Neural Networks with Optimizations

“…For instance, the introduction of siloxane groups leads to increased mechanical flexibility and gas permeability, while carboxylic acid esters alter the hydrophobicity and thermal stability of polymeric structures. , Cross-linking with electronically saturated aliphatic scaffolds allows polymers with increased thermal stability and solubility. On the basis of these features, a diverse library of 1200 oligomeric structures with unknown chirality was computationally generated and structurally characterized, using the valence-aware polarizable reactive force field (ReaxPQ-v) methoda recently introduced variant of the polarizable reactive force field (ReaxPQ) method. ,− Modifications on the polymeric subset were made to ensure covering a broad range of variable functionalities. The building blocks included organic fragments having electron-withdrawing (e.g., CF 3 -, Cl-, F-), electron-donating (e.g., OMe-, tBu-, Me-), amide-type group substituents (NH–R–CO), sterically hindered groups, and unsaturated methylene moieties.…”