Figure 3.

Transfer learning (TL) for λ of polymers using 19 observations. (a) The upper left plot shows 19 observed properties against predicted values given by directly trained random forests. The other panels present the prediction performance of transferred random forests trained using neural network features acquired from prelearning on C_V (small molecules), and the viscosity, ρ, T_g, and T_m of polymers. The predicted and fitted values in the 5-fold CV are colored orange and blue, respectively. (b) Scatter plot matrix of observed properties in PoLyInfo for T_g (°C), T_m (°C), ρ (g/cm³), viscosity (η, dL/g) in log scale, C_P (cal/g °C), and λ (W/mK). (c) Prediction performance of transferred random forests trained using neural network features acquired from prelearning of the ionization energy (E_ion), n, cohesive energy (E_coh), Hildebrand solubility parameter (δ), and electronic dielectric constant (ϵ_e) in Polymer Genome.