Page 232 - Read Online

Page 10 of 13 Yuan et al. J. Mater. Inf. 2026, 6, 17

Figure 6. Performance of the incremental learning framework on the sample molecules across multiple metal substrates. (A) Visualization
of molecular detection results for 4,4’-di(pyridin-4-yl)-1,1’-biphenyl in STM images acquired from Au(111), Ag(111), and Cu(111) surfaces.
All STM images span 20 nm × 20 nm. Bounding boxes highlight the successfully identified molecules. In the model visualization, yellow
and orange represent gold (Au) and copper (Cu) atoms, respectively; (B) Quantitative performance comparison across substrates using
mAP@0.5, showing the model’s ability to maintain high detection accuracy (0.81, 0.95, and 0.99) under varying imaging conditions. STM:
Scanning tunneling microscopy.

As shown in Figure 6, the model demonstrates robust molecular recognition even when the substrate type
changes, which typically introduces differences in background texture, surface electronic structure, and
imaging noise. Despite these challenges, the incremental learning approach enables the model to retain the
discriminative features of the molecule while adapting to substrate-induced variations. This robustness is
largely attributed to the fact that the incremental training strategy does not simply overwrite previously
learned weights with new data, but instead uses a balanced training approach in which data from both old
and new substrates are sampled in each training epoch, with the old dataset comprising 10% of the training
data. This design allows the model to maintain equilibrium between previously learned features and new
information, resulting in stable recognition performance under diverse imaging conditions.

Our experiments confirm that even when Cu(111) images present higher noise levels and lower contrast, the
model is still able to accurately detect and classify the molecule based on its essential structural features.
Furthermore, variations in substrate properties not only affect image noise and contrast but can also
influence the self-assembly behavior of molecules on the surface. For instance,
4,4’-di(pyridin-4-yl)-1,1’-biphenyl molecules may exhibit slightly different arrangements on Au(111),
Ag(111), and Cu(111) due to differences in adsorption energy and intermolecular interactions, which in turn
affect molecular packing density and orientation [33,34] . By incorporating data from multiple substrates, the
incremental learning model learns to recognize key molecular structures and features regardless of such
variations.

Lastly, we note an additional advantage of the incremental learning framework: improved training efficiency.
Traditional approaches typically require training a separate model for each substrate, which is
computationally costly and time-consuming. In contrast, our incremental approach dynamically updates a