Asynchronous Cross-Modal Dynamic Graph Learning for Intelligent Sensing of AI Computing Infrastructure Expansion.

The rapid expansion of artificial intelligence computing infrastructure and the semiconductor industry has made the dynamic sensing of policy planning, technological innovation, capital investment, and physical construction essential for industrial monitoring and resource allocation. However, policy documents, remote-sensing imagery, patent relations, and economic indicators exhibit substantial heterogeneity, temporal misalignment, and variations in data quality, making it difficult for conventional unimodal or synchronously fused methods to accurately characterize the continuous evolution of construction activities. To address these challenges, a multimodal industrial sensing dataset for artificial intelligence computing infrastructure was constructed, and an artificial intelligence-driven asynchronous cross-modal dynamic graph prediction framework was developed. Differential temporal lags among policy announcements, capital investment, patent growth, and physical construction are learned through an asynchronous cross-modal alignment module. Interference caused by cloud contamination, missing text, patent disclosure delays, and incomplete economic observations is dynamically suppressed through a reliability-aware fusion mechanism. Time-varying propagation relationships among countries, enterprises, patents, and industrial parks are further modeled using an economically modulated dynamic heterogeneous graph. Experimental results demonstrate that, in next-window construction event prediction, the proposed model achieved an Accuracy of 91.72%, a Precision of 91.08%, a Recall of 90.64%, an F1-score of 90.86%, and a ROC-AUC of 95.67%, significantly outperforming baseline methods including XLM-R, Swin Transformer, TimesNet, HGT, and TGN. In construction intensity forecasting, the model achieved an MAE of 0.104, an RMSE of 0.153, a MAPE of 10.91%, an R2 of 0.854, and a Pearson correlation coefficient of 0.928. Ablation experiments further confirmed the effectiveness of asynchronous alignment, reliability calibration, missing-modality compensation, economic modulation, and long-term graph memory. The proposed approach provides a temporally interpretable intelligent sensing method for monitoring computing center construction, evaluating industrial park expansion, allocating energy and communication infrastructure, and planning equipment supply.