Digital twins (DTs) have emerged as powerful tools for improving measurement accuracy and uncertainty estimation in metrology systems. However, ensuring high precision in metrology DTs requires accurate error source identification in the virtual-to-physical (V2P) process, as well as closed-loop parameter updates in the physical-to-virtual (P2V) process. Therefore, this study proposes a machine-learning-based error source identification framework for metrology DTs, enabling data-driven calibration and uncertainty control. By leveraging simulated measurement data from the virtual entity, a transformer-based machine learning model is developed to identify the error sources and predict the error values. The identification results can facilitate V2P decision-making, such as the recalibration of error source parameters to ensure the required uncertainty level. The proposed method is validated through two case studies on a virtual coordinate measuring machine (CMM), where different measurement paths are used to evaluate model performance. The experimental results demonstrate that the proposed method can effectively identify various error types and accurately predict error values, showing its potential to improve the reliability and accuracy of metrology DTs through efficient V2P decision-making. • A general machine-learning framework is proposed for error source identification in metrology digital twin systems. • The method uses simulated data from virtual measurement systems to support data-driven calibration. • A transformer-based model enables accurate classification and estimation of systematic geometric errors. • The approach enhances virtual-to-physical decision-making and improves uncertainty control in digital metrology.