@Article{cmc.2026.081876,
AUTHOR = {Yuxuan Feng, Lili Liang, Guanglu Sun, Yanrui Wei},
TITLE = {MSA-ConvNeXt: Predicting Magnetism of Doped Two-Dimensional Nanomaterials via Multi-Scale Convolution and Attention Mechanisms},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/cmc/online/detail/27160},
ISSN = {1546-2226},
ABSTRACT = {In doped two-dimensional nanomaterials, magnetism is one of the important physical properties. By introducing foreign doping atoms or molecules, the electronic structure of the material can be effectively regulated, leading to changes in magnetic behavior. Currently, magnetic property prediction has achieved considerable results with the help of traditional CNNs, but there are still obvious limitations: (1) The feature extraction of dopant sites is constrained by fixed receptive fields, making it difficult to characterize local structural perturbations in the vicinity of dopant atoms and their spatial influence propagating to surrounding regions; (2) CNNs lack the capability to model long-range dependencies between non-neighboring atoms and their chemical bonds, thereby weakening the representation of long-range interactions within the material. In this study, we propose Multi-Scale and Attention ConvNeXt (MSA-ConvNeXt) based on multi-scale convolution and attention mechanisms, which consists of the following two core modules: (1) The Multi-scale Convolution Attention Block (MCAB), which models local structural perturbations around dopant atoms and their spatial effects via parallel multi-scale convolutions. It uses a serial channel and spatial attention mechanism to adaptively recalibrate multi-scale features, highlighting the response of doping related regions and enhancing the ability to express dopant-site information; (2) The Visual Geometry Group–Swin Transformer (VGG-Swin) architecture extracts structural features of dopant sites using VGG convolutions to prevent the attenuation of structural information during global relationship modeling. Subsequently, the Swin Transformer is introduced, which uses the self-attention mechanism to dynamically weight and globally associate features at different spatial locations, in order to depict the long-range correlations between non-neighboring atoms and their chemical bonds with the dopant-site. Experiments conducted on a doped two-dimensional nanomaterial dataset constructed from the CMR database demonstrate that the proposed model outperforms existing methods in terms of accuracy and F1-score. Specifically, MSA-ConvNeXt achieves an accuracy of 91.66%, representing an improvement of 1.65% over the next best model. In addition, all experimental results are averaged over multiple independent runs (with five different random seeds), demonstrating the stability and reliability of the model’s performance. Ablation studies further validate the effectiveness of each module design.},
DOI = {10.32604/cmc.2026.081876}
}