Toward predictive additive manufacturing: ML-based hardness prediction for Ti–6A1–4 V in laser powder bed fusion

Abstract

Laser powder bed fusion (LPBF) of Ti–6Al–4 V frequently yields high hardness, but the result is sensitive to coupled process settings. This work investigates whether hardness can be predicted directly from the process parameters to support data-driven process planning. A literature-derived dataset of 136 cases was compiled with laser power, scan speed, layer thickness, hatch distance, and spot size as inputs and Vickers hardness as the output. Five regression models were compared: gradient boosting, random forest, support vector regression (SVR), a feed-forward neural network, and L1-regularized linear regression. The models were assessed using cross-validation and a data split of 85%/15% for training and testing. Ensemble tree methods delivered the most accurate predictions, with gradient boosting achieving the highest explanatory power \(\left({R}^{2}=0.705\right)\) with low error (RMSE = 22.015 HV; MAE = 13.873 HV; MAPE = 3.834%), followed by random forest (\({R}^{2}=0.595\)). The neural network, linear model, and SVR performed substantially worse due to limited capacity or misspecification for the interaction-rich small-sample setting. Distributional analysis showed that the dataset clusters around common LPBF presets (e.g., 30-μm layers, ∼0.1-mm hatch), while hardness was concentrated near ∼380 HV. Overall, the results indicate that ensemble models capture the dominant nonlinearities linking LPBF parameters to hardness and can serve as practical surrogates for predictive design and preliminary optimization of Ti–6Al–4 V builds.