부산대학교 도서관

The primary bottlenecks faced by the laser powder bed fusion (LPBF) process is the identification of optimal process parameters to obtain high density (> 99.8%) and a good surface finish (< 10 µm) in the fabricated components. Prediction of optimal process maps with the help of machine learning (ML) models is still challenging due to extensive training data, which proves to be expensive in additive manufacturing. In view of this, the present study employs six different supervised ML algorithms on a comparatively small data set of 33 experiments to predict relative density and surface roughness. It has been observed that input data (predictor) curation can increase the accuracy of the ML models even with a small data set. In the ML prediction model, the mean absolute percentage error (MAPE) was reduced by 30% (relative density) and 21.94% (surface roughness) with volumetric energy density as an input parameter instead of laser power, scanning speed, hatch space, and layer thickness. The choice of non-dimensional energy input as a universal predictor allows for an increase in training size and the translation capability of trained ML models from one machine/material combination to another. The ML model based on increased training data size (198 for relative density and 173 for surface roughness) procured from the material processed/fabricated on different LPBF machines showcased reasonable R2 values of 79.11% and 80.3% for relative density and surface roughness, respectively.