Leveraging simulated and empirical data-driven insight to supervised-learning for porosity prediction in laser metal deposition

Abstract

The advent of digital-twin manufacturing in additive manufacturing (AM) is to integrate the physical world of real-time 3D printing with the digital world of a simulated print. This paper contributes to digital-twin manufacturing in laser-based additive manufacturing by combining melt pools’ simulated thermal behavior via finite element analysis (FEA) and melt pools’ empirical thermal behavior via pyrometry-based sensors. Studying the thermal behavior of melt pools based on heat transfer characteristics determines melt pools’ porosity and part quality. FEA uses Godak’s moving heat flux to capture the melt pools’ physically bound temperature profile in three dimensions. Simulated data helps to mitigate the influence of measuring errors from real-world data and provides non-observable data such as gradient changes of thermal behavior at the curvature of the 3D melt pool. The pyrometer captures empirical temperature behavior, including uncertainty and randomness introduced to the process. A significant knowledge gap exists when predicting melt pool porosity accurately with theoretical FEA and empirical in situ evidence alone. The gap is bridged by combining the data sources, specifically, feature engineering via functional principal component analysis (empirical data source) and capturing the melt pool's 3-D temperature shape profile via FEA (simulated data source). A hybrid model predicts melt pool porosity by capturing the strengths of prior simulated and posterior in situ empirical data by matching simulated melt pools to real-world empirical melt pools. Moreover, comparing predicted porosity labels with true porosity labels of Ti–6Al–4V thin-wall structure from laser metal deposition verified the proposed interpretable and robust supervised-learning model's validity. This methodology can apply to other materials and part shapes printed under various additive-manufactured printers.