Novel Machine Learning-Driven Platforms for In-Situ Prediction of Vertical and Top Surface Roughness in Laser Powder-Bed Fusion

Abstract

Controlling and optimizing surface roughness remains a significant challenge in Laser Powder Bed Fusion (LPBF), as roughness profoundly influences fatigue life, mechanical performance, and post-processing (e.g., machining) costs. While in-situ monitoring has emerged as a key approach for real-time defect detection, predicting surface roughness, particularly for both top and vertical surfaces of parts being printed, remains underexplored. Existing studies predominantly rely on camera-based methods, which usually suffer from the limitations such as lack of viewability of vertical surfaces covered by loose powder particles in LPBF, sensitivity to ambient light, resolution constraints, and the need for additional optical equipment. This research pioneers a novel photodiode-based in-situ monitoring framework integrated with machine learning (ML) algorithms to predict surface roughness in real-time for both top and vertical surfaces of LPBF-printed parts. For top surface roughness prediction, the methodology involves capturing light intensity signals from the melt pool using an on-axial photodiode, incorporating additional process parameters, and training multiple ML models to predict surface roughness at a fine spatial resolution (690 µm × 510 µm), including edges and corners. The framework is rigorously evaluated across a wide range of roughness values, demonstrating its robustness in adapting to process parameter variations. For vertical surface roughness prediction, this study introduces the first-ever in-situ framework using photodiode signals to overcome challenges posed by loose powder coverage, which obstructs conventional sensing techniques. Key time-domain and frequency-domain features are extracted from photodiode signals captured near vertical surfaces and combined with essential process parameters to train ML models. Among the five ML models evaluated, Random Forest (RF) and eXtreme Gradient Boosting (XGB) demonstrated the highest accuracy and lowest error rates. Incorporating in-situ data significantly improved RF’s performance, increasing R² from 0.35 (using process parameters alone) to 0.78, confirming the effectiveness of this approach. This research introduces an innovative pathway for real-time surface roughness prediction in LPBF, enabling enhanced quality assurance, process optimization, and defect mitigation. The integration of photodiode signals with advanced ML algorithms enables precise, on-the-fly assessment of both top and vertical surfaces, enhancing the ability to detect and address irregularities as they occur. By addressing the limitations of traditional camera-based methods, this photodiode-ML framework provides a fast, adaptive, and scalable solution for real-time surface monitoring, paving the way for more advanced quality control strategies, and promising greater reliability and consistency in the production of high-performance components.