Page 230 - Read Online

Page 8 of 13 Yuan et al. J. Mater. Inf. 2026, 6, 17

Figure 5. STM molecular images and prediction results. (A) Chemical structures of six molecular categories (M1-M6); (B) STM images
and prediction results obtained using the well-trained model. The detected molecules are highlighted with color-coded markers: blue
squares denote molecule M1, orange squares denote M2, green squares denote M3, grey squares denote M4, cyan squares denote M5,
and purple squares denote M6. STM: Scanning tunneling microscopy.

demonstrating the effectiveness of the replay mechanism in mitigating catastrophic forgetting while
maintaining the model performance across multiple categories. These results indicate that the training
strategy effectively preserves previously learned knowledge, as reflected by the minimal performance
degradation.

Figure 5 provides a more intuitive demonstration of the effectiveness of the incremental learning approach.
These results are based on the model after five rounds of incremental learning, where the trained weights
integrate the learned representations of six molecular categories (M1 to M6). Figure 5A displays six
representative molecular categories (M1-M6), each corresponding to a specific precursor or coordination
unit used in STM studies of on-surface molecular self-assembly, i.e., M1: 2,4,6-tri(pyridin-
4-yl)-1,3,5-triazine; M2: 4,4’-(5’-(4-(pyridin-4-yl)phenyl)-[1,1’:3’,1’’-terphenyl]-4,4’’-
diyl)dipyridine; M3: 4,4’-di(pyridin-4-yl)-1,1’-biphenyl; M4: hexakis[4-(4’-pyridyl)phenyl]
benzene; M5: 4,4’-(1,3-phenylene)bis-pyridine; and M6: 3,3’,5,5’-tetra(pyridin-4-yl)-1,1’-biphenyl.