*Result*: Coordinated physics-integrated machine learning and its application in tunnel engineering.

*Accurate and reliable modeling of control parameters is essential for the intelligent operation of large-scale engineering equipment. However, the complexity of these physical systems, combined with the scarcity and noise contamination of engineering data, makes it challenging for data-driven artificial intelligence methods to satisfy the comprehensive modeling requirements of such high-risk scenarios. To address these issues, this study proposes a coordinated physics-integrated machine learning framework comprising three components: a knowledge-driven module, a data-driven module, and an integration layer. The knowledge-driven module employs mechanical principle and dimensional analysis theory to extract the underlying logic governing the equipment system. The integration layer serves to incorporate the knowledge-driven and data-driven modules while resolving their coordinated challenges. By tailoring the learning strategy, it endows the data-driven module with global sensitivity and degrees of freedom aligned with the underlying logic, thereby reinforcing the physical guidance and constraints of the knowledge-driven module and enabling the algorithm to suppress noise responses while preserving efficient, stable learning of essential patterns. The proposed method is applied to the thrust modeling of tunnel boring machines. Ablation studies based on real-world tunneling data elucidate the operational mechanisms of each module within the proposed method. Compared to conventional machine learning approaches, this method reduces mean relative error by at least 60.8 %, significantly mitigates training bias due to distribution shifts, and demonstrates superior stability under perturbations in training data. • A novel physics-integrated machine learning is proposed for harsh data scenarios. • Knowledge-driven module is designed to extract physical logic from complex system. • Integration layer is introduced to match physical logic and machine learning. • Two specific integration-layer schemes designed for diverse application scenarios. • Proposed method is validated in tunnel boring machine thrust modeling tasks. [ABSTRACT FROM AUTHOR]*