Page 110 - Read Online

Shang et al. J. Mater. Inf. 2025, 5, 52 https://dx.doi.org/10.20517/jmi.2025.36 Page 3 of 15

Additive exPlanations (SHAP) method, we provide an in-depth analysis of the key factors that lead to a
wide range of values for the work function of MXenes, thereby transforming the traditional “black box”
model into a transparent “glass box” model. Our approach not only provides valuable insights into the
determinants of the work function but also offers a scalable strategy applicable to a broader range of
material categories.

MATERIALS AND METHODS
The data containing 4,034 materials was extracted from Computational 2D Materials Database (C2DB) [25,26] .
Ninety-eight characteristics were recorded for each material, including properties such as Fermi energy
level, elastic modulus, material ID, volume, and relative molecular mass. To validate the feasibility of our
[27]
stacking model approach, we conducted tests using various models, including random forest (RF) ,
gradient boosting decision tree (GBDT) , and lightGBM (LGB) , as meta-models. To enhance the
[28]
[29]
[30]
model’s predictive capability, we generated additional effective features using the SISSO method and
incorporated them into the training process. Furthermore, we employed SHAP values to analyze the
[31]
importance of each feature, thereby gaining deeper insights into the model’s decision-making mechanism.
All data processing and ML tasks were conducted in Python, with data processing handled by the Pandas
library and ML algorithms implemented using the Scikit-learn library. The performance of the models was
evaluated by using MAE and coefficient of determination (R ). These metrics are mathematically defined in
2
the Supplementary Equations (1) and (2).

RESULTS AND DISCUSSION
Workflow
As illustrated in Figure 1, the workflow begins with rigorous data selection and cleaning for model training
and validation. Herein, the raw data of the work function of MXenes were curated from the recently created
C2DB. Then, the SISSO algorithm was utilized to construct the descriptors with enriched physics insights
between features and target properties, which will effectively improve the accuracy and interpretability of
ML. Meanwhile, a stacked model that combines predictions from various base models into a refined meta-
model was constructed, which was confirmed to possess not only enhanced generalization but also
effectively reduced errors, overfitting, and biases. Additionally, SHAP was leveraged to determine feature
importance, providing nuanced insights into the role of atomic-layer compositions and feature
interdependencies in governing work function behavior. Ultimately, ML, combined with material science
knowledge, provides a deeper understanding of the factors influencing the work function of MXenes.

Data preparation and feature screening
Our study focused on a subset of 275 MXenes that were screened from 4,034 candidate materials,
particularly the work function values obtained through DFT calculations [32,33] . For the purpose of model
development, 80% of the MXene data were allocated to the training set, while the remaining 20% were
reserved for testing. This dataset provides a solid foundation for in-depth analysis and predictive modeling
due to its richness. However, the 98 features contained in the dataset also pose challenges for modeling,
prompting us to pay attention to feature engineering. Feature engineering, as an important aspect of ML,
directly affects the performance of ML models. For instance, the excessive feature set will lead to the well-
[34]
known dimensionality curse , which not only amplifies the computational demands and resource usage of
the model but also exacerbates the risk of overfitting. Overfitting is a common phenomenon during the
training of ML models. It refers to the model excessively learning the details and noise of the training data,
to the extent that it fails to capture underlying patterns. This leads to a significant decline in the model’s
performance on new data [24,35] . Therefore, it is crucial to control overfitting and ensure the model has good
generalization ability. Here, we proposed a new metric to quantify the degree of model overfitting; that is,