CGformer: AI Overcomes "Myopia" in Material Design

Artificial intelligence is revolutionizing materials science by dramatically accelerating the traditionally slow and complex process of discovering new materials. What once took decades can now be achieved in days or even hours, enabling faster translation of cutting-edge materials from lab to industry. This includes high-performance polymers, corrosion- and heat-resistant composites, high-efficiency perovskite solar materials, and even the long-sought goal of room-temperature superconductors. At the heart of this transformation are AI models that predict material properties, with crystal graph neural networks (GNNs) such as CGCNN and ALIGNN becoming foundational tools in this domain. However, despite their success, these conventional GNNs suffer from a fundamental architectural limitation: they rely solely on local, neighbor-based information exchange. In these models, each atom can only communicate with its immediate neighboring atoms, resulting in a restricted "field of view" that prevents the capture of long-range atomic interactions. These long-range effects—crucial for determining properties like ionic conductivity, thermal stability, and mechanical strength—are effectively invisible to the model. This "myopia" limits predictive accuracy and has been a key bottleneck in AI-driven materials discovery, often leading to failed experimental validations despite promising computational predictions. To overcome this critical challenge, researchers from the Artificial Intelligence and Microstructure Lab (AIMS-Lab) at Shanghai Jiao Tong University have developed CGformer—a novel AI architecture that introduces global attention mechanisms inspired by Transformers to break the limitations of traditional crystal GNNs. CGformer enables true "global perception" of crystal structures by allowing every atom to interact directly with all others, regardless of distance. The architecture operates through a multi-stage process. First, the 3D crystal structure is converted into a graph, where atoms are nodes and chemical bonds are edges, each enriched with physical and chemical features. The core innovation lies in the global attention module, which replaces the step-by-step local message passing with simultaneous, long-range information exchange—transforming the model from "neighborly whispers" into a "full-field broadcast." To enhance precision, CGformer incorporates several specialized encodings: centrality encoding to assess each atom’s structural importance, spatial encoding to capture actual atomic distances, and edge encoding to integrate bond-specific properties. This enables the model to detect complex, long-range cooperative effects that govern material behavior. The effectiveness of CGformer was rigorously tested using a particularly challenging case: high-entropy sodium-ion solid electrolytes (HE-NSEs). These materials exhibit extreme structural disorder and their performance depends heavily on long-range interactions—making them an ideal benchmark for evaluating global perception capabilities. In comparative experiments, CGformer outperformed established models like CGCNN, ALIGNN, and SchNet. On a dataset of 2,300 sodium-ion conductors, CGformer reduced the mean absolute error (MAE) by 25.7% on the training set and nearly 10% on the test set. Even more impressively, when applied to the data-scarce HE-NSE problem, CGformer achieved high accuracy using only 238 newly computed samples after pre-training on a large general dataset—demonstrating exceptional small-sample learning capability with a cross-validation MAE of just 0.0361. The algorithm successfully screened over 140,000 candidates and identified six promising HE-NSEs, all of which were experimentally synthesized and validated. The resulting materials exhibited room-temperature ionic conductivities as high as 0.256 mS/cm—significantly outperforming baseline materials. Published in the journal Matter under the title "CGformer: Transformer-Enhanced Crystal Graph Network with Global Attention for Material Property Prediction," the work was led by Prof. Jinjin Li and Prof. Fuqiang Huang from AIMS-Lab at Shanghai Jiao Tong University, with Ph.D. students KeHao Tao and Jiacong Li as co-first authors. This breakthrough exemplifies a new research paradigm: solving real-world scientific challenges not by applying existing AI tools, but by innovating the underlying algorithmic framework. By embedding physical insights—such as the importance of long-range interactions—into the model design, CGformer achieves both higher accuracy and deeper scientific interpretability. The framework is highly transferable and holds promise for diverse applications, including lithium-ion conductors, thermoelectric materials, and photocatalysts. As part of a broader mission, AIMS-Lab has developed over 100 AI-driven software tools such as AlphaMat and AphlaBio, and pioneered ManuDrive, the world’s first time-dimension AI industrial brain system, enabling intelligent manufacturing with minimal data and full domestic control. Through a closed-loop ecosystem spanning algorithm development, platform innovation, and industrial deployment, AIMS-Lab continues to bridge the gap between fundamental research and real-world impact. The lab welcomes researchers passionate about advancing AI for science and industry.