Despite the effectiveness of most graph-based representation methods in capturing the crystal geometry characteristics, they fall short in intuitively describing phenomena such as ionic transport behavior which are often determined by the atom-unoccupied regions in the mobile sublattice (namely interstitial network). Here, we develop a Structure Divide-and-Conquer Graph Representation method based on graph neural network (SDCGNNdk), for unveiling structure-activity relationships of transport barriers by incorporating Domain Knowledge (e.g., site energy information, thresholds for ion accessibility, etc.), where crystal geometry and interstitial network topology are combined to construct a dual-structure crystal graph. For driving the proposed model, we construct a graph-based dataset for the prediction of activation energy (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">${E}_{a}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mrow> <mi>E</mi> </mrow> <mrow> <mi>a</mi> </mrow> </msub> </math></EquationSource> </InlineEquation>), i.e., the energy barrier hindering ionic transport, covering over 18,000 ionic compounds from the Inorganic Crystal Structure Database (ICSD), including Li+, Na+, K+, Ag+, Cu(2,3)+, Mg2+, Zn2+, Ca2+, Al3+, F−, and O2−. SDCGNNdk achieves high prediction performance of <InlineEquation ID="IEq2"> <EquationSource Format="TEX">${E}_{{\rm{a}}}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mrow> <mi>E</mi> </mrow> <mrow> <mi mathvariant="normal">a</mi> </mrow> </msub> </math></EquationSource> </InlineEquation> with R2 of 91.30%, outperforming conventional GNNs by more than 20% on average and offering insights into structure-activity relationships by quantifying the contributions of crystal geometry and interstitial network characteristics to transport barriers. This work provides an accurate graph representation and GNN framework, demonstrating potential for extension to predicting other properties relevant to interstitial network of inorganic compounds.

Structure divide-and-conquer: dual graph representation for accurate ionic transport barrier prediction of inorganic compounds

期刊论文
出版模式：

开放获取
发表日期：

2026年4月3日

查看全文

Zhengwei Yang,
Linhan Wu,
Bing He,
Maxim Avdeev,
Siqi Shi,
Yue Liu

摘要

Despite the effectiveness of most graph-based representation methods in capturing the crystal geometry characteristics, they fall short in intuitively describing phenomena such as ionic transport behavior which are often determined by the atom-unoccupied regions in the mobile sublattice (namely interstitial network). Here, we develop a Structure Divide-and-Conquer Graph Representation method based on graph neural network (SDCGNN_dk), for unveiling structure-activity relationships of transport barriers by incorporating Domain Knowledge (e.g., site energy information, thresholds for ion accessibility, etc.), where crystal geometry and interstitial network topology are combined to construct a dual-structure crystal graph. For driving the proposed model, we construct a graph-based dataset for the prediction of activation energy ( ${E}_{a}$ $E_{a}$ ), i.e., the energy barrier hindering ionic transport, covering over 18,000 ionic compounds from the Inorganic Crystal Structure Database (ICSD), including Li⁺, Na⁺, K⁺, Ag⁺, Cu^(2,3)+, Mg²⁺, Zn²⁺, Ca²⁺, Al³⁺, F⁻, and O²⁻. SDCGNN_dk achieves high prediction performance of ${E}_{{\rm{a}}}$ $E_{a}$ with R² of 91.30%, outperforming conventional GNNs by more than 20% on average and offering insights into structure-activity relationships by quantifying the contributions of crystal geometry and interstitial network characteristics to transport barriers. This work provides an accurate graph representation and GNN framework, demonstrating potential for extension to predicting other properties relevant to interstitial network of inorganic compounds.