This study proposes an energy-saving solution for the high-energy-consuming gas-lift catalyst system in semiconductor manufacturing, integrating intelligent control and thermal energy recovery technologies to address the energy demands and carbon emission challenges posed by AI chip fabrication. By introducing digital twin technology and the Gradient Boosting Regression (GBR) model, three sub-models (GBR-Cone, GBR-Energy, GBR-Temp) were developed to predict and optimize system airflow, energy consumption, and temperature differential. Using actual operational data for modeling and cross-validation, the models achieved a prediction accuracy of over R² = 0.95, demonstrating high stability and generalization capability. The results show that under simulator control, the heat recovery rate increased from 68.7% to 98.5%, and daily heater power consumption dropped from 6,018 kWh to 214 kWh, achieving a power-saving rate of 96.4%. Upon full-scale implementation across the plant, the annual energy savings reached 6.91 million kWh, with a carbon reduction of 3,270 tons of CO₂e, highlighting significant economic and environmental benefits. To enhance system resilience, the study also introduced feature drift detection and model switching mechanisms. In cases of sensor anomalies or deviations, the system can automatically switch to the temperature differential model for calibration, ensuring control accuracy and stability. This research not only provides a practical and feasible energy-saving technology but also establishes a replicable intelligent control framework, offering critical value for the semiconductor industry's transition toward sustainable development. Future applications may extend to other high-energy-consuming process equipment and integrate ESG and carbon inventory mechanisms to drive comprehensive low-carbon transformation across the industry.
Keywords: Intelligent Control, Digital Twin-Based Thermal Energy Recovery, Gas-Lift Catalyst System, Machine Learning Models