Page 226 - Read Online

Page 4 of 13 Yuan et al. J. Mater. Inf. 2026, 6, 17

ID, which is then used in the “Dataset Generation” step. The generated dataset serves as the foundation for
model training, ensuring that all relevant categories are included. Additionally, when generating the dataset,
users can enable the incremental learning feature to create a more generalized dataset. In the incremental
learning phase, the model is updated with newly added data without complete retraining from scratch,
significantly enhancing its adaptability, especially as new types of species and patterns are discovered in the
experimental image data.

After dataset creation, the “Train & Detect” step utilizes the YOLOv9 model to perform both object detection
and instance segmentation, identifying key features in the image data. YOLOv9, the latest version of the
YOLO family of models , combines advanced detection algorithms with high efficiency, enabling accurate
[29]
and fast object detection even for complex datasets. It improves upon its predecessors by incorporating more
robust backbone networks, enhanced anchor-free mechanisms, and dynamic label assignment strategies,
making it particularly suitable for detecting small-scale molecular features in high-resolution images. Once
detection is completed, the results are processed by the “Data Processing” module, which extracts valuable
insights such as the identified classes and the average area of detected instances. These insights are crucial for
understanding surface nanostructures, molecular self-assembly, and the distribution of reactants and
products. The seamless integration of incremental learning ensures that the model evolves continuously as
new data become available, maintaining accuracy as experimental conditions change and allowing
researchers to adapt their analysis to emerging surface phenomena. This combination of YOLOv9’s advanced
detection capabilities and incremental learning provides a powerful framework for analyzing molecular
systems.

During training, we found that hyperparameters have a significant impact on the model’s performance.
Improper hyperparameter configurations can drastically reduce training efficiency, potentially wasting
several hours without yielding satisfactory results. In contrast, an appropriate combination of
hyperparameters can accelerate convergence, improve model stability, and significantly enhance overall
performance. This effect becomes particularly evident in deep learning tasks, where training is
computationally expensive. A poorly tuned learning rate or batch size can either cause divergence or lead to
excessively slow convergence, while suboptimal augmentation or regularization parameters may result in
underfitting or overfitting, making the training process inefficient and unstable. To address this challenge, it
is crucial to adopt a systematic and automated hyperparameter optimization strategy that goes beyond
manual trial-and-error. Given the large number of hyperparameters, manually adjusting each value is
extremely time-consuming and practically impossible. Therefore, we employed a Bayesian optimization
algorithm to search for the optimal combination of hyperparameters. Compared with conventional methods
such as grid search and random search, Bayesian optimization can find near-optimal solutions with fewer
iterations, especially in high-dimensional hyperparameter spaces. It builds a surrogate model [e.g., Gaussian
process (GP)] to predict the behavior of the objective function and leverages an acquisition function [such as
expected improvement (EI), probability of improvement, or upper confidence bound] to balance exploration
of unknown regions and exploitation of promising areas, thereby avoiding inefficient or blind searches.
Unlike grid search, which exhaustively evaluates every point in a predefined parameter grid, or random
search, which samples configurations uniformly at random, Bayesian optimization intelligently selects the
next point to evaluate based on prior knowledge from past observations. This approach is particularly
efficient when each model evaluation is computationally expensive, as in our training pipeline, where a single
training run can take several hours.

The surrogate model we utilized, GP, offers a probabilistic interpretation of the function being optimized. GP
estimates not only the mean performance for any hyperparameter configuration but also the associated
uncertainty, enabling the acquisition function to determine whether to explore new regions or exploit areas