Page 33 - IJAMD-2-1

International Journal of AI for
Materials and Design
ML molecular modeling of Ru: A KAN approach

A B

C D

Figure 3. Distribution patterns before and after normalization. (A) Original energy values of the dataset samples; (B) normalized energy values displaying
preserved distribution pattern; (C) original force features in the PC1-PC2 space; and (D) normalized force features demonstrating maintained relative
relationships. The consistent patterns between pre- and post-normalization data validate the effectiveness of the standardization process.

A B C

Figure 4. Scattering plots of (A) training dataset, (B) validation dataset, and (C) test dataset, demonstrating consistent cross-shaped patterns across all
three subsets.

regularization (λ = 0.02) and entropy regularization The model employs a sequential architecture comprising
43
(λ_entropy = 4.0) to enhance the model’s ability to three distinct layers designed for regression analysis. The
32
generalize. This is crucial for preventing overfitting and input layer accepts three features, corresponding to the
maintaining robust performance on new, unseen data. reduced-dimensionality data obtained through PCA of
atomic structures. These inputs are processed through
2.5. CalHousNet feedforward neural network model a hidden layer containing three nodes, where each node
(CalHousNet) implements a linear transformation followed by a ReLU

We implemented CalHousNet, a fully connected activation function. This activation introduces essential
45
feedforward neural network using the PyTorch framework. non-linearity into the model, enabling the capture
44
Volume 2 Issue 1 (2025) 27 doi: 10.36922/ijamd.8291