1.

Schematic workflow of the ECW-TL framework. (1) A baseline model is trained using an active-learning workflow. (2) A diverse subset of configurations is selected from the baseline data set. (3) DFET/ECW calculations are performed on the selected configurations. (4) The model is fine-tuned using ECW data, during which part of the neural network (e.g., the embedding network, highlighted by the red dashed circle) is frozen to avoid overfitting. (5) MD simulations are then carried out to assess convergence. If convergence is not achieved, additional data is selected and the fine-tuning cycle is repeated.