Machine Learning Approaches for Predicting Microstructure in PBF-LB Ti-6Al-4V Components

Sammanfattning

Laser Powder Bed Fusion (PBF-LB) offers unrivalled geometric freedom for Ti-6Al-4V parts, yet the steep thermal gradients inherent to the process make the resulting microstructure, and thus mechanical performance, difficult to predict. This thesis critically reviews peer-reviewed studies published between 2010 and 2025 in which machine-learning (ML) models are trained to forecast PBF-LB microstructural features. Eight empirical or simulation investigations satisfy stringent inclusion criteria, collectively spanning input spaces that range from nominal process parameters and physics-based thermal descriptors to in-situ sensor streams and image-derived metrics. Although well-tuned models reach median internal accuracies around R² ≈ 0.90, no single algorithm family dominates once dataset size and validation rigour are normalised. Three systemic obstacles remain: models built on single-machine data lose on average ΔR² ≈ 0.17 when transferred to new equipment; complex architectures frequently obscure underlying process–structure relationships; and credible prediction intervals are reported in only a minority of cases, limiting regulatory acceptability. Deep-learning hybrids and real-time sensor fusion emerge as promising directions. However, they currently demand data volumes that few laboratories can supply. Progress therefore hinges on assembling federated, physics-informed datasets: harmonised, multi-lab repositories that embed heat-transfer priors and make external validation routine. Such resources would enable reproducible benchmarking, support uncertainty-aware optimisation and accelerate the transition toward closed-loop, microstructure-controlled additive manufacturing.