In-situ residual stress mitigation of laser powder bed fusion Ti-6.5Al-2Zr-1Mo-1V using an active learning framework

Titanium (Ti) alloys fabricated by laser powder bed fusion (LPBF) often suffer from high residual stress (RS), which in turn increases manufacturing time and cost due to the need for post-processing treatment. In this study, an active learning framework, demonstrated using Ti‑6.5Al‑2Zr‑1Mo‑1 V (TA15) alloy, was developed to simultaneously minimize RS and deformation while ensuring low porosity. Seven machine learning algorithms were evaluated systematically, and among them, gradient-boosted decision trees, LightGBM, and eXtreme gradient boosting were found to give the most accurate predictions for RS, porosity, and deformation, respectively. Then, the optimized models were coupled with Non‑dominated Sorting Genetic Algorithm III and an active learning framework. This combined approach led to a notable drop in the Z reduction rate to 3.4% after four iterations. Microstructure of the as‑built (AB) sample using the optimization parameter (OP) exhibited finer α′-martensitic laths (0.619 ± 0.012 vs. 0.704 ± 0.009 μm), lower kernel average misorientation (0.210 ± 0.004 vs. 0.450 ± 0.006), and more homogeneous stress distributions compared with the sample fabricated using the machine‑recommended parameter (MP). This resulted in an enhanced mechanical performance, namely the OP-AB sample achieving a tensile strength of 1315.8 ± 30.5 MPa, an elongation of 7.7% ± 0.2%, an impact toughness of 27.2 ± 0.7 J/cm², and a fracture toughness of 64.7 ± 1.5 MPa m^1/2, matching or surpassing those of the stress-relieved sample fabricated using the MP. Furthermore, an aerospace backplate was fabricated using the OP parameter, and the result indicated that a 43.1% reduction in maximum RS and a 62.8% decrease in distortion were achieved. Taken together, this work establishes a data-driven optimization strategy for LPBF-ed Ti alloys that achieves in-situ stress mitigation without post-processing, facilitating broader industrial adoption of the proposed data-driven optimization strategy.