AI agent-driven virtual in-situ calibration for intelligent building digital twins

Building digital twins (BDTs) couple real-time data with virtual models to enhance automated operations yet maintaining the accuracy of physical sensors and virtual models is still challenging. To address this challenge, this study introduces an artificial intelligence (AI) agent–driven virtual in-situ calibration (VIC) that integrates the Brick schema ontology. It can automate the full calibration process and make results interpretable. The AI agent coordinates three toolkits including benchmark modeling, correction modeling, and problem solving to automate the full calibration process. It can develop the benchmark model, get expression form and input variable for the correction model, and optimize the calibration target with minimal human intervention based on prompts. To validate the automated performance, the method is applied to the heating, ventilation, and air conditioning (HVAC) system. The root mean square error (RMSE) decreases from 1.10 °C to 0.30 °C for the return chilled water temperature, from 0.97 m³/h to 0.088 m³/h for the chilled mass flow. Meantime, the RMSE of the return chilled water temperature virtual model can reduce from 1.13 °C to 0.49 °C, confirming calibration accuracy gains alongside higher automation. The Brick based integration explicitly links calibration targets, models, and building components, enabling consistent knowledge representation and interpretable calibration reports. Overall, the study contributes a novel, generalizable framework that combines AI agents, VIC, and Brick for reliable and scalable BDT implementations.