Multi-objective optimization of the laser powder-bed fusion process of 17–4 PH stainless steel: Balancing quality and processing rate

• A new physics-statistics-informed MOO framework was built to optimize LPBF process. • The correlation between surface roughness and density was built by an MOGP model. • An EHI-based active learning approach was proposed for reducing MOO computation time. • An interface for LPBF multi-objective optimization was developed for end-users. Reducing build time while maintaining fabrication quality is vital to broadening the industrial adoption of laser power-bed fusion (LPBF) additive manufacturing. However, the processing parameters that maximize processing rate often conflict with those optimized for mechanical properties. This study integrated physical and statistical models to establish the relationship between these optimization objectives and the process parameters, and thus utilized a multi-objective optimization (MOO) approach to maximize the localized processing rate while maintaining high part-level quality. The physical model, built based on Eagar-Tsai’s solution, offers time-efficient predictions of the melt pool dimensions and localized processing rates. The statistical model, developed using a multiple-output Gaussian process (MOGP), incorporates the correlation between the part-level objectives based on their experimental measurements. Compared with the other three traditional MOO algorithms in terms of Pareto front solution diversity, root mean square error (RMSE), and run time, we found that: (i) the nonlinear relationships between the part-level objectives, i.e. surface roughness and relative density, could be established using the MOGP; (ii) the MOGP model demonstrated higher prediction accuracy for part-level objectives than both the second-order polynomial (SOP) and standard Gaussian process (GP) models; and (iii) the developed MOO, based on the expected hypervolume improvement (EHI) active learning strategy, proved superior to those established by non-dominated sorting genetic algorithm II (NSGA-II), achieving a 14 % reduction in RMSE while reducing the run time by 25 %.