Page 231 - Read Online

Yuan et al. J. Mater. Inf. 2026, 6, 17 Page 9 of 13

Using the trained weights, a single prediction was performed, and the results shown in Figure 5B
demonstrate clear and accurate identification of the corresponding molecular structures. The predicted
patterns align well with the molecular structures in Figure 5A, indicating that the model effectively captures
and retains the features of all six categories without significant performance degradation. To further quantify
this performance, the evaluation metrics indicate a precision of 0.959, recall of 0.931, mAP@0.5 of 0.976, and
mAP@0.5:0.95 of 0.792, confirming the model’s robustness across classes. Specifically, the per-class
mAP@0.5:0.95 values for M1 through M6 are 0.698, 0.728, 0.685, 0.794, 0.904, and 0.945, respectively. These
results highlight the model’s strong generalization capability and its ability to maintain high accuracy and
localization precision across diverse molecular categories, even after multiple stages of incremental updates.

These findings further highlight the strength of the incremental learning strategy, which enables the model to
continuously learn new molecular categories while retaining knowledge of previously learned ones. The
predictions indicate that the model can handle the complexity of molecular assemblies and accurately detect
and classify molecules, thus illustrating its robustness and generalization capability after incremental
learning. The clear patterns and accurate bounding boxes in the predictions reflect the model’s ability to
integrate diverse molecular features, making it a powerful tool for analyzing high-resolution images of
complex molecular arrangements.

Batch effects arising from different STM imaging environments, such as variations in tunneling current
stability, tip geometry, or detector sensitivity, may introduce distributional bias into the image data. To
mitigate this issue, the dataset includes STM images collected under diverse experimental conditions on
Au(111), Ag(111), and Cu(111) substrates. By incrementally introducing data from these different imaging
environments while replaying earlier samples, the model continuously calibrates itself across multiple
acquisition domains, thereby reducing the risk of bias. In addition to the experiments discussed above, the
model’s performance was further evaluated on images of the same molecular species acquired on three
different substrates, namely Au(111), Ag(111), and Cu(111) [Figure 6]. The results show that the incremental
learning strategy not only allows the model to continuously adapt to new data but also exhibits strong
generalization across diverse imaging conditions. Despite variations in surface structures, noise levels, and
contrast among different substrates, the model successfully identified the target molecule with no noticeable
degradation in detection accuracy. This outcome indicates that the replay mechanism effectively preserves
the essential features of the learned molecule, enabling robust recognition across heterogeneous datasets. The
ability to adapt to different substrates highlights the strong transferability and generalization capability of the
incremental learning framework. In STM imaging, the substrate type plays a critical role in determining the
overall image characteristics. For example, Au(111), Ag(111), and Cu(111) surfaces differ in lattice constants
and electronic structures, leading to significant variations in background textures, contrast levels, and noise
patterns in the resulting images. Conventional deep learning detection models are often sensitive to such
differences; when trained exclusively on data from a single substrate, their performance typically declines
significantly when tested on another . To overcome this issue, the aforementioned incremental learning
[32]
strategy combined with a replay mechanism was adopted.

In our experiments, we progressively introduced STM images of the target molecule obtained on Au(111),
Ag(111), and Cu(111). At the initial stage, the model was trained solely on Au(111) data and achieved high
detection accuracy. We then incrementally incorporated images from Ag(111) and Cu(111) substrates while
simultaneously replaying Au(111) images to reinforce the retention of previously acquired knowledge. This
strategy ensured that the model maintained strong recognition performance for Au(111) molecules while
adapting to the unique imaging characteristics of Ag(111) and Cu(111). The final evaluation results
demonstrate that the model maintains consistently high accuracy across all three substrates (with only minor
variations in mAP@0.5), thereby validating the effectiveness of the incremental learning approach.