Linking process parameters to residual stress and distortion in directed energy deposition repair via machine learning and response surface methodology

• FE-driven ML framework established for optimising multi-layer L-DED filling repair. • Upscaling strategy integrates geometric constraints for consistent fill volume. • MLP model outperforms peers with superior accuracy (R² = 0.98, NRMSE = 2.7%). • RSM identifies moderate energy density as optimal balance for stress & distortion. • Scan speed exerts dominant influence on repair quality indicators. Laser-directed energy deposition (L-DED) filling repair is often compromised by printing defects like cracking induced by excessive thermal residual stresses. Optimising process parameters is challenging, as complex thermal histories make trial-and-error costly and simulations computationally inefficient. To bridge this, a finite element (FE)-driven machine learning (ML) framework was developed to optimise multi-layer multi-track filling repair bulk quality. The methodology ensures consistent fill volume by enforcing nominal single-track dimensions. A thermomechanically validated FE model generated training data via design of experiment (DoE) strategies. Among evaluated algorithms, the multilayer perceptron (MLP) achieved superior accuracy (R 2 = 0.98, NRMSE = 2.7%) as an efficient surrogate. Integrated response surface methodology (RSM) highlighted a critical trade-off, revealing moderate energy density as the optimal compromise for balancing residual stress and distortion. Furthermore, parameter influence analysis identified scan speed as the dominant control variable, closely followed by laser power and preheat temperature. Ultimately, this framework provides a robust and efficient tool for defining optimal process windows in L-DED filling repair.