PhyAgent-TWM: Reproducibility package for physics-grounded agentic tool-wear management
Year: 2026
Type: software
Abstract: PhyAgent-TWM is a fully reproducible research package accompanying the study From Prognosis to Action: A Physics-Grounded Agentic AI Framework for Autonomous and Uncertainty-Aware Tool-Wear Management in Smart Manufacturing. The framework integrates multimodal condition evidence, physics-grounded degradation modelling, remaining useful life (RUL) prognosis, tool-group-aware conformal uncertainty quantification, explicit engineering knowledge, sequential decision logic, and deterministic safety constraints for tool-management actions. The repository provides the complete reproducibility benchmark used for decision-level validation. It contains a deterministic physics-grounded digital surrogate comprising 300 independent tool trajectories and 32,382 machining cycles, partitioned at tool level into 180 training, 60 calibration, and 60 held-out test trajectories. The benchmark includes process-context variables, sensor-derived features, wear and RUL targets, controlled sensor disturbances and dropout events, fixed data partitions, leakage controls, and reproducible evaluation outputs. The package contains the synthetic benchmark dataset, data dictionary, fixed train/calibration/test splits, experiment configuration, predictive and decision metrics, robustness and ablation results, sequential replay outputs, figure-generation assets, manuscript figures, supplementary material, repository metadata, licensing information, and documentation required to reproduce the reported analyses. PhyAgent-TWM is intended to evaluate the transition from prognostic prediction to auditable maintenance action without claiming causal machine-control effects that are unsupported by existing public milling datasets. Public resources including the Piecuch-Zabinski milling dataset, PHM 2010 milling challenge, NASA Milling Wear repository, and the Denkena-Klemme-Stiehl cross-machine dataset are therefore treated as external-validation targets rather than redistributed data sources. The repository supports transparent replication, benchmarking, extension of the PhyAgent-TWM architecture, and future external validation of physics-grounded, uncertainty-aware and safety-constrained autonomous maintenance systems.
Language: English
Subfield: Safety, Risk, Reliability and Quality
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213435646