Abstract
<title>Abstract</title> <p>Achieving zero-defect production in five-axis machining of aerospace titanium alloys remains difficult because process dynamics change continuously as tools wear, thermal fields evolve, and vibration modes shift with tool orientation. Conventional digital twins, built once from calibrated physics models or offline-trained data models, degrade in fidelity as these conditions drift away from the calibration point, and by the time a defect appears in the finished part it is too late to intervene. This paper proposes a self-evolving digital twin architecture that unifies physics-informed machine learning, deep reinforcement learning, and multi-agent coordination into a single closed-loop system capable of predicting defects before they occur and adjusting machining parameters in real time. The digital twin embeds governing equations of cutting mechanics, heat conduction, and tool wear directly into the loss function of a graph-structured neural network, so that the surrogate model remains physically consistent even when sensor coverage is sparse. A Bayesian model-updating layer continuously re-estimates uncertain physical parameters (friction coefficient, thermal softening exponent, wear rate constant) from streaming force, vibration, acoustic emission, and temperature signals, allowing the twin to evolve its internal representation without full retraining. On top of this twin, a soft actor-critic reinforcement learning agent, constrained by physics-derived safety filters, searches the feasible space of feed rate, spindle speed, depth of cut, and tool axis inclination to jointly minimize surface roughness deviation, tool wear rate, specific cutting energy, and cycle time. A three-agent coordination layer (quality agent, energy agent, and productivity agent) negotiates a Pareto-consistent operating point at each control interval using a weighted Nash bargaining rule, resolving conflicts that a single scalarized objective cannot represent. The framework is evaluated on a numerical case study of five-axis ball-end milling of Ti-6Al-4V, calibrated against published cutting-force, temperature, and tool-wear data for this alloy. Compared against five baseline strategies (fixed Taguchi-optimized parameters, classical model predictive control, a pure data-driven LSTM predictor, single-agent deep Q-network control, and a physics-only finite element digital twin without learning), the proposed framework reduces surface roughness variance by 41.3%, cuts tool wear rate by 27.8%, lowers specific cutting energy by 19.6%, and reduces defect escape rate from 6.4% to 0.7% over a simulated 500-part production campaign, with all improvements significant at the 0.01 level under paired t-tests and Wilcoxon signed-rank tests. The paper details the mathematical formulation, algorithmic implementation, simulation protocol across ANSYS, MATLAB, Python, and Siemens NX, and discusses industrial deployment implications, limitations, and directions for experimental validation.</p>