Data-Centric Machine Learning for Surface Roughness in Laser Powder Bed Fusion

Understanding and tailoring surface roughness on Laser Powder Bed Fusion (L-PBF) components is an ongoing challenge in key aerospace and medical applications. Large experiments or computationally intense simulations to understand the phenomena are often needed. However, machine learning (ML) is an alternative approach that can accelerate and enhance experimental and simulation-based efforts. Effective ML models rely on large, informative and diverse training data to predict across different domains. There is a scarcity in the development and availability of such datasets. This study proposes and executes a database approach to build a large training dataset that predicts surface roughness and generate synthetic surfaces on ideal CAD profiles. A modular and extensible relational data model was developed and implemented to ingest data from a wide range of builds and specimens. The database currently spans data from over 25 different builds and surface roughness data from over 1000 specimens. An ML workflow was developed using the database wherein models are trained with the goal of generalizability in predicting surface roughness for unseen processing conditions. The preliminary results show significant potential for achieving robust prediction of surface roughness (balanced accuracy ~90% in challenging test sets). This work aims to address two challenges affecting L-PBF reliability: (i) A stronger understanding of L-PBF surface roughness and (ii) improved generalizability and usability of ML models training for L-BPF applications.