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

Area of Science:

• Artificial Intelligence
• Semiconductor Industry
• Industrial Monitoring
• Data Fusion

Background:

• Rapid expansion of AI computing infrastructure necessitates dynamic sensing of policy, technology, investment, and construction.
• Heterogeneous, temporally misaligned, and variable-quality data from policy, remote sensing, patents, and economic indicators challenge conventional monitoring methods.
• Accurate characterization of continuous construction evolution is difficult with existing unimodal or synchronously fused approaches.

Purpose of the Study:

• To develop an AI-driven asynchronous cross-modal dynamic graph prediction framework for industrial sensing.
• To construct a multimodal industrial sensing dataset for AI computing infrastructure.
• To address challenges in monitoring AI infrastructure construction due to data heterogeneity and temporal misalignment.

Main Methods:

• Developed a multimodal industrial sensing dataset for AI computing infrastructure.
• Constructed an AI-driven asynchronous cross-modal dynamic graph prediction framework.
• Employed an asynchronous cross-modal alignment module to learn differential temporal lags, a reliability-aware fusion mechanism to suppress interference, and an economically modulated dynamic heterogeneous graph to model propagation relationships.

Main Results:

• Achieved 91.72% Accuracy, 91.08% Precision, 90.64% Recall, 90.86% F1-score, and 95.67% ROC-AUC in next-window construction event prediction.
• Obtained MAE of 0.104, RMSE of 0.153, MAPE of 10.91%, R2 of 0.854, and Pearson correlation of 0.928 in construction intensity forecasting.
• Ablation experiments validated the effectiveness of asynchronous alignment, reliability calibration, missing-modality compensation, economic modulation, and long-term graph memory.

Conclusions:

• The proposed AI-driven asynchronous cross-modal dynamic graph prediction framework effectively monitors AI computing infrastructure construction.
• The approach provides a temporally interpretable intelligent sensing method for industrial park expansion, resource allocation, and supply chain planning.
• Demonstrated significant performance improvements over baseline methods in both event prediction and intensity forecasting.