This paper presents a multiscale modelling system based on a digital twin that can predict defects in metal additive manufacturing in real time, with primary validation scoped to Laser Powder Bed Fusion (LPBF). While the framework is architected to generalise across powder-bed and directed-energy deposition (DED) processes, all experimental evaluations are conducted on LPBF using the publicly available NIST AM-Bench benchmark dataset of IN625 and Ti-6Al-4 V specimens. It combines microscale melt pool dynamics with mesoscale thermal fields and macroscale structural deformation via hierarchical physics-informed neural network (PINN) surrogates, with a real-time Internet of Things (IoT) backbone of sensors. It has a three-tier architecture of edge, fog, and cloud, with inference distributed across latency-sensitive edge nodes, fog-level surrogate aggregation, and cloud-based digital twin calibration. Thermal imaging, acoustic emission, and optical monitoring provide sensor-based feedback to the numerical model, which can be used to adaptively control the process in real time during Laser powder bed fusion (LPBF) and directed energy deposition (DED). The hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) is an extractor of spatio-temporal defect signatures used to classify six defect classification categories, including porosity, lack-of-fusion, cracking, balling, keyholing, and delamination, with a macro-averaged F1-score of 0.9841. Formation of the multiscale coupling, Bayesian calibration and surrogate training formalised in twenty-five governing equations. It was experimentally verified on the publicly available NIST AM-Bench dataset that the proposed DT-MSM structure achieves a mean defect-detection rate of 98.72% and an inference latency of 11.3ms per frame on edge hardware, outperforming seven other baselines. The framework achieves reductions in scrap rate (34.6 per cent) and predictive maintenance lead time (41.2 per cent) in a simulated LPBF production scenario, with improved predictive ability for porosity and cracking compared with offline simulations. Mean plus standard deviation of the results is reported across five random seeds, over which the statistical significance is verified using a paired t-test (p < 0.01). The proposed methodology supports smart manufacturing and quality control in the new-generation production systems for metal additive manufacturing.
Alfattani et al. (Sun,) studied this question.