Fig. 2: Transfer learning curves for Hydrogen on Copper surface and for the Ti-Al-V alloy systems, trained based on the “small” foundation model.
From: Fine-tuning foundation models of materials interatomic potentials with frozen transfer learning
For the Cu-H2 system, root mean squared errors (RMSEs) of energies, force components and force components of the H atoms only are shown in panels (a–c), respectively. For the Ti-Al-V alloy system, root mean squared errors (RMSEs) of energies, force components and virial stress components are shown in panels (d–f), respectively. The points correspond to the percentages of the respective datasets, namely 2, 5, 10, 20, 40, 60, 80 and 100% for both systems. The layers were frozen correspondingly to f6 (pink circles), f5 (yellow diamonds), f4 (green triangles), and f0 (blue circles). Grey squares mark the learning curve of the from-scratch-trained MACE model in case of the Cu-H2 system, and the grey dashed line marks the errors of the custom ACE model in case of the Ti-Al-V alloy system.
