Uncertainty-aware deep ensembles for real-time porosity prediction in metal additive manufacturing with calibrated confidence for operator decision support

Metal additive manufacturing (AM) in-situ monitoring systems achieve high defect detection accuracy but generally provide deterministic predictions without confidence estimation, limiting their applicability for safety-critical decisions. This paper presents an uncertainty-aware deep ensemble framework that integrates multimodal in-situ sensing, including thermal imaging, acoustic emission, and spectral monitoring, for layer-wise porosity and lack-of-fusion prediction. A feature-level fusion architecture based on a shared CNN–LSTM backbone performs simultaneous porosity regression with prediction intervals and defect classification into nominal, gas porosity, and lack-of-fusion categories. Predictive uncertainty is quantified by decomposing aleatoric and epistemic components using a law-of-total-variance formulation over ensemble predictive distributions, followed by calibration using temperature and variance scaling. Physics-guided augmentation, cross-modal synchronization constraints, stochastic modality dropout, and regime-aware MixUp improve ensemble diversity and robustness against distribution shifts. An uncertainty-gated cost-sensitive decision strategy maps predictive distributions to optimal manufacturing actions, including continue, inspect, re-scan, and abort. Evaluation on NIST AMMT, AM-Bench 2022, and multi-material datasets (Ti-6Al-4V and Inconel 718) with micro-CT validation demonstrates improved calibration, defect prediction reliability, and reduced intervention cost compared with deterministic approaches. The proposed framework connects deep-learning-based AM monitoring with quantitative non-destructive evaluation reliability principles.