Transfer Learning for Quality Optimization in Laser and Electron Beam Powder Bed Fusion

Transfer learning targets quality defects in metal additive manufacturing

Metal additive manufacturing technologies face significant hurdles with consistent part quality, as defects like lack of fusion and keyhole porosity emerge from suboptimal process parameters, Preprints.org reported in its review of laser and electron beam powder bed fusion processes.

Traditional parameter optimization relies on expensive and time-consuming experimental campaigns that are strictly material-specific. Machine learning models offer a solution by mapping relationships between process parameters and defects, but these algorithms require large datasets that do not exist for newly developed alloys.

Transfer learning addresses this data scarcity by reusing knowledge from well-characterized alloys to improve performance in related domains. Instead of training models from scratch, the approach adapts pre-trained systems using limited amounts of new data.

Optical and acoustic monitoring systems track melt pool dynamics

Process monitoring captures real-time data without requiring post-built inspections, according to ScienceDirect. Optical and acoustic monitoring systems serve as the two primary methods for tracking these builds.

Optical systems measure visible lights and thermal signals emitted from the powder bed and melt pool using digital cameras and photodiodes. Photodiodes offer higher sampling rates and lower costs compared to camera imaging, though digital cameras provide detailed surface visuals.

Acoustic monitoring captures sound waves, atmospheric pressure fluctuations, and plasma density variations through microphones.

Machine learning models classify defects across multi-machine production

Supervised learning remains the most widely used machine learning technique for classifying powder bed anomalies, ScienceDirect noted. Researchers have deployed convolutional neural networks and semantic segmentation to categorize powder bed conditions and predict ultimate tensile strength.

Industrialized multi-machine production setups face distinct hurdles regarding machine-to-machine variability. Variations in shielding gas flows, laser spot characteristics, and scan strategies across different printers mean that optimized parameters on one machine do not guarantee identical outputs on another.

MDPI reported that transferring process models between different printers requires limited training data, ranging between 7 and 49 data vectors on target machines for Ti-6Al-4V datasets. Researchers have utilized pre-trained models on synthetic porosity data to calibrate actual experimental prints of AlSi10Mg0.5.

Future parameter development relies on small-dataset adaptation

Balancing defect suppression with high productivity remains a central requirement for industrial scale adoption. Utilizing knowledge transfer across diverse printer hardware bridges the gap between extensive laboratory datasets and practical workshop deployment.

Fine-tuning parameters to improve laser powder bed fusion repeatability

Optimizing processing results requires fine-tuning parameters like preheating temperature, laser scan speed and power, or layer height, alongside characteristics of the laser spot such as diameter, focus or caustic, and scanning strategies.

Despite promising advantages, poor process repeatability and part consistency remain major challenges hindering broader adoption of laser powder bed fusion parts.