Accurate prediction of the solidified microstructure in laser powder bed fusion (LPBF)-processed components is critical because heterogeneity and spatial anisotropy in the solidified microstructure lead to variation in functional properties. This paper presents a physics-aware, real-time, data-integrated machine learning approach for predicting the solidified microstructure of LPBF-processed Inconel 718 parts. Currently, the prediction of solidified microstructure in LPBF is realized primarily through three approaches: (i) empirical modeling; (ii) blackbox modeling using in-situ sensor data; and (iii) whitebox modeling leveraging physics-based process simulations. Each of these approaches has inherent limitations. Empirical and blackbox models lack generalizability because they do not account for geometry-dependent causal thermal-fluid phenomena governing microstructural evolution. Physics-based whitebox models are computationally demanding and overlook the stochasticity inherent to the process. To overcome the foregoing limitations, this work establishes a graybox modeling approach. This approach combines temperature fields predicted by a physics-based thermal model with real-time data acquired from in situ infrared thermal imaging and optical tomography sensors. The graybox model is trained to predict the following microstructural aspects of LPBF-processed Inconel 718 parts: melt pool depth; primary dendritic arm spacing; crystallographic texture, orientation, and grain aspect ratio; and microhardness. The graybox model predicted the solidified microstructure with an accuracy approaching 95% (R²). By contrast, the prediction accuracy of blackbox and physics-based whitebox models ranged between R² ~ 60% and 85%. Thus, this work takes a critical step towards a rapid, accurate, in-situ, and non-destructive Born Qualified assessment of LPBF part quality. • Physics-aware machine learning (graybox modeling) approach for predicting solidified microstructure in LPBF-processed Inconel 718 parts. • Approach integrates lightweight computational thermal modeling and real-time, in-situ multi-sensor data. • Approach validated on Inconel 718 parts (12 parts) with varying processing conditions and design conditions. • Four aspects of solidified microstructure were predicted - meltpool depth, dendritic spacing, grain structure (aspect ratio, orientation, grain size, and texture), and microhardness • Graybox modeling approach fusing sensor signatures and simulation-derived thermal metrics predicts with a statistical accuracy exceeding 90% (R 2 ) • Purely data-driven machine learning (blackbox modeling) and physics-based modeling (whitebox modeling) have predictability in the range of 60% - 85% (R 2 )
Deshmukh et al. (Sun,) studied this question.