This research investigates the performance of different transfer learning strategies in the context of materials science, specifically focusing on CrCoNiFeMn alloys. The study compares the effectiveness of three strategies: Freeze, Progressive, and Multi-Task, analyzing their impact on Mean Absolute Error (MAE) and performance gap. The methodology involves varying the amount of target data and evaluating the strategies' success rates and average improvements over baseline.
Key findings
Best transfer achieved with CrCoNiFe using Strategy B (Progressive) at 100% data with an MAE of 0.0185 eV/atom.
One-third of transfers achieve a 5% gap at 50% data.
Strategy B (Progressive) is identified as the best strategy with an average improvement of 5.0%.
Limitations & open questions
Limited to specific alloys and may not generalize to other materials.
Further research needed to validate findings across a broader range of materials and conditions.