Machine Learning-Driven Optimization of Laser Powder Bed Fusion: From Powder Characterization to Process Parameters Tailoring for Ti6Al4V Powders

Abstract

The quality of parts produced by laser powder bed fusion (LPBF) is influenced by a complex interplay of factors, including process parameters, feedstock powder behavior, melt pool conditions, and environmental variables such as gas flow direction. Developing a comprehensive quality assurance system for LPBF necessitates incorporating these diverse factors into the analysis. Despite their significance, the intrinsic characteristics of the powder—particularly powder bed quality—have rarely been the primary focus of global research. Key attributes such as particle size, morphology, and rheology are crucial in determining powder bed quality and laser-powder interaction. Therefore, accurately estimating powder rheology or flowability before printing is crucial for achieving consistent, high-quality production. This thesis explores the use of machine learning (ML) algorithms to predict powder flowability based on simpler characteristics such as size, morphological features, and packing, without resorting to costly and time-consuming experiments. It focuses on a sample set of standard and off-size Ti6Al4V (Ti64) powders, with two main objectives: 1) assess ML regressors and equation fitting for flowability prediction, using the small original dataset, and 2) overcome data limitations’ impact on regression through data augmentation techniques. Despite the limitations of the original dataset, the support vector regressor (SVR) delivered excellent performance in repeated random cross-validation (CV) analysis for the dynamic rheology characteristic of Specific Energy (SE), achieving a mean absolute percentage error (MAPE) of approximately 2.6% across six random CV iterations. Furthermore, the application of genetic programming's symbolic regression significantly enhanced the estimation accuracy of rheological characteristics, specifically the unconfined yield strength at 3 kPa (UYS3) and 9 kPa (UYS9), using the original dataset. The Synthetic minority over-sampling technique (SMOTE) proved effective in achieving admissible regression results for characteristics such as basic flow energy (BFE) and Hall flowability, whereas other explored methods of simple ML models training, and GP-SR equation-fitting had been unsuccessful. After employing machine learning for powder characterization, it is crucial to integrate the identified powder characteristics into the optimization of LPBF process parameters. From an application perspective, selecting the appropriate set of process parameters is often influenced by the size distribution range of the powder used and the specific properties of the target part to be optimized. Therefore, incorporating the powder size range into systems developed for tailoring and optimizing process parameters is highly beneficial. Consequently, this thesis aimed to develop a comprehensive system for optimizing LPBF process parameters by integrating photodiode-based melt pool monitoring (light intensity signals) with ML. The proposed approach employs regression algorithms to determine optimal process parameters, including laser power, scan velocity, hatch distance, and energy density. These parameters are tailored based on the size range of the Ti64 powder, melt pool/light intensity signatures, and the desired properties of the final part, such as density, hardness, and surface roughness. Three different regressor algorithms of feed-forward neural network (FFN), random forest (RF), and extreme gradient boosting (XGBoost) were optimized and trained for the proposed task. Comparison analyses were conducted between algorithms in terms of regression accuracy and generalization capabilities. It was concluded that while all three algorithms could achieve satisfactory results on this dataset, the RF algorithm exhibited higher overall accuracy and better generalization. RF yielded an overall average MAPE of ~2% (R2 of ~94%) in unseen samples analysis.