Session: 04-22-01: Machine Learning-Based Modeling, Prediction, and Optimization of Advanced Manufacturing and Materials System with Multiphysical Phenomena

Paper Number: 166383

Data Driven Prediction and Optimization of Laser Powder Bed Fusion Melt Pool Characteristics 

Laser Powder Bed Fusion (LPBF) is a widely utilized metal additive manufacturing (AM) process known for its ability to fabricate intricate geometries with high mechanical strength. However, achieving defect-free parts remains a challenge due to complex thermal dynamics and process variability. The quality of LPBF-fabricated components is primarily determined by the melt pool morphology, which depends on key process parameters, including laser power, scan speed, and layer thickness. Improper selection of these parameters can lead to defects such as porosity (keyhole and lack of fusion), balling, and residual stresses, compromising the structural integrity of the final component. Thus, optimizing LPBF parameters is crucial but challenging due to the multi-scale and multi-physics nature of the process, which traditionally relies on costly and time-intensive experimental trials. This study proposes a data-driven approach utilizing machine learning (ML) models to predict and optimize LPBF melt pool characteristics, thereby reducing reliance on trial-and-error experimentation.

A dataset comprising 1,499 simulated single-track laser scans was generated using an Ansys Thermo-Mechanical solver. The melt pool geometries were categorized into four distinct groups based on their aspect ratios and fusion characteristics: desirable melt pools (Green), under-melting defects (Yellow), balling defects (Orange), and keyhole porosity (Red). Statistical analysis, including Analysis of Variance (ANOVA) and Tukey’s Honest Significant Difference (HSD) tests, was performed to assess the significance of laser power, scan speed, and layer thickness in determining melt pool quality. The results indicate that laser power has the strongest influence on melt pool formation, exhibiting a strong correlation (0.71) with defect generation. Higher scan speeds tend to produce more favorable melt pools, whereas excessive laser power at low scan speeds leads to deep keyhole defects.

To predict and classify melt pools efficiently, several ML models were implemented, including logistic regression, Naïve Bayes, decision trees, ensemble learning techniques (Random Forest, Extra Trees, Gradient Boosting), and fully connected neural networks (FNNs). The Weighted Neural Network (WNN) with an adjusted class-loss function demonstrated the highest performance, achieving an overall F1-score of 0.97 and a perfect recall (1.0) for optimal melt pools (Green category). This underscores the effectiveness of machine learning in accurately classifying LPBF melt pools, thereby enabling rapid identification of ideal process parameters.

The findings suggest that machine learning can serve as a powerful tool for optimizing LPBF manufacturing processes and reducing material waste, defects, and computational expenses associated with high-fidelity simulations. However, real-world validation through experimental LPBF trials remains necessary to confirm the reliability of the ML-driven predictions. Future research should focus on integrating in-situ monitoring data, reinforcement learning techniques, and computational efficiency analysis to enhance predictive accuracy and real-time process control. Additionally, multi-objective optimization approaches can be explored to balance print quality, mechanical properties, and processing speed for diverse alloys and part geometries.

In conclusion, this study demonstrates that machine learning provides an efficient, data-driven methodology for LPBF process optimization, offering a significant advantage over traditional trial-and-error approaches. By leveraging ML models, manufacturers can accelerate the development of high-quality metal AM components, benefiting industries such as aerospace, automotive, and biomedical engineering where precision and reliability are paramount.

Presenting Author: Mohammad Akram University of New Haven

Presenting Author Biography: Mohammad Akram
Department of Mechanical and Industrial Engineering
University of New Haven
300 Boston Post Rd, West Haven, CT 06516

Mohammad Akram is currently pursuing his Ph.D. at the University of New Haven, where his research focuses on polymer extrusion processes enhanced through the application of machine learning and artificial intelligence. He completed his Bachelor's in Engineering from Jadavpur University, where he built a strong foundation in materials and mechanical sciences, which now supports his innovative research at the intersection of polymers, AI, and process engineering.

Authors:

Debajyoti Adak Indian Institute of Technology, Kharagpur
Mohammad Akram University of New Haven
Somnath Roy Indian Institute of Technology, Kharagpur
Ganesh Balasubramanian University of New Haven