Fig. 2: Influence of PT size and FT strategy on FT model performance.
From: Optimal pre-train/fine-tune strategies for accurate material property predictions
Percentage change in test R2 scores of PT-FT models with respect to corresponding scratch models, with the PT done on a FE and b BG. The FT datasets are indicated along the x-axis, while the FT dataset size is kept to 800. PT dataset sizes are indicated with different colored bars. c, d Circular bar plots illustrating the performance in terms of c R2 scores and d 1-MAEs for select PT-FT models. Different FT strategies are illustrated by the different colored bars. The range of the bars is indicated by the text annotations provided across the concentric circles in both panels.
