Ensuring defect‐free parts in metal additive manufacturing (AM) is vital for safety‐critical components, yet it often relies on costly trial‐and‐error and slow computed‐tomography (CT) inspections. Here, we introduce a two‐stage, machine‐learning‐driven quality‐assurance framework for laser powder‐bed fusion (L‐PBF) that balances predictive modeling with rapid, vibration‐based, non‐destructive evaluation (NDE).

Stage 1: Process‐Parameter Optimization
A full‐factorial design of experiments (DoE) covering laser power, scan speed, and hatch spacing (75 successful builds) feeds a polynomial Ridge‐regression surrogate. Using nested 5‐fold cross‐validation and grid‐search tuning, the final quadratic Ridge model achieved a validation R² of 0.51 (RMSE ≈ 0.59 % density), capturing just over half of the variance in out‐of‐sample relative‐density measurements .

Stage 2: Vibration‐Based Defect Screening
Specimens are subjected to modal excitation and frequency‐response analysis (150–390 kHz), yielding 10,000 interpolated features. After in‐fold LARS feature‐selection and stability‐thresholding, three classifiers (5‐NN, SVC, and MLP) were evaluated via nested CV. The best model (MLP) attained 0.81 ± 0.08 accuracy, with true‐negative rates above 90%, but modest true‐positive recall (25–35%).

Integrated Impact
By combining proactive parameter tuning with vibration‐based NDE, the framework enables the removal of the majority of defective builds before certification and replaces hour‐long CT scans with minute‐scale vibration tests. This dual‐stream approach lays the groundwork for scalable, in‐situ quality assurance in AM, offering a path toward the real‐time monitoring and digital‐twin certification of complex parts.