Integration of physics-based models and machine learning approaches for tool wear prediction in machining

Hybrid physics-data models are increasingly used for tool-wear prediction in turning and milling, yet cross-study comparison and industrial adoption remain limited by heterogeneous reporting of targets, sensing chains, validation protocols, and deployment assumptions. This review synthesizes 122 Web of Science studies published from 2020 to 2025 and contributes: (i) a P0 - P5 taxonomy comprising a data-only baseline and five physics-data integration patterns; (ii) a cross-study evidence matrix consolidating performance ranges, matched-protocol baseline gains, and validation-tier composition; and (iii) a minimum reporting checklist mapped to architectural, validation, and intervention boundaries. Among the 51 studies with aggregable matched-protocol comparisons, the median relative error reduction over auditable baselines is 44.3%. Physically grounded descriptors are the most prevalent pattern, representing 39.3% of the corpus and showing the highest median gain, 60.5%, followed by mechanistic-state and residual architectures, with 17.2% prevalence and 59.8% median gain. These advantages remain conditional: within-dataset partitioning dominates, 59.0%; explicit out-of-distribution testing appears in only 31.1%; and cross-machine validation is absent. Deployment and sustainability evidence are also weakly reported: drift detection is absent from the online subset, timing characterization is incomplete in nearly half of online-positioned studies, and sustainability accounting boundaries are fully declared in only 3.3% of the corpus. The central finding is that architectural maturity has outpaced reporting maturity. The proposed taxonomy, evidence matrix, and checklist provide a reproducible basis for interpreting hybrid model claims and translating monitoring-grade results into deployment-grade evidence.