PSI - Issue 83

136 C. Mallor et al. / Procedia Structural Integrity 83 (2026) 130–137 parameters and levels, as ideally, both should be the same. The simulated deflections, ௜ , vs. the estimated deflections from the twin-cantilever surrogate model, ௜ ௌெ , are shown in Fig. 10.

Fig. 10. Deflection tip results simulated ௜ vs. ROM predicted values ௜ ௌெ for the different DOE tests. When the surrogate predictions were compared with the original FEM values at the trained points, the agreement is excellent, the error band ranged from about -0.17% to 0.24%. In other words, the ROM reproduced the behaviour of the validated full-order model almost exactly inside the training domain. An additional check was interpolation at temperatures not seen during training. Additional FEM runs were therefore performed at 110 and 300 ºC while retaining the same power and scan-speed levels. At these intermediate temperatures, the surrogate still performed well, with prediction errors between roughly -1.2% and 0.24%. If required by the application, additional simulations can be performed to improve the accuracy of the ROM by generating an expanded database. From a practical view, the surrogate provides a speed advantage over the full-order model, while remaining close to the 3D printing physics. 4. Conclusions This paper presents an integrated benchmark study for PBF-LB/M-induced distortion in a metal twin-cantilever. The work aims to be a single methodology including experiments, thermo-mechanical finite element modelling, and reduced-order surrogate development to create a physics-based prediction chain for vertical deflection. The benchmark results show that the Abaqus framework reproduces the measured thermal history and the released deflection response accurately. The study also shows that build-plate preheating is the dominant parameter for reducing distortion in the investigated process window, whereas the effects of power and scan speed are comparatively less dominant. On top of the validated full-order model, the Twinkle TRD framework produced an accurate and interpretable surrogate model. The ROM reproduced trained FEM points and remained accurate when interpolating at intermediate temperatures. In practical terms, it provides the low computational cost needed for future decision-support and digital twin uses without breaking the link to the governing process physics. Future work should extend the same workflow to additional geometries and materials; couple the surrogate layer to broader distortion-compensation or process-calibration routines; and examine how the framework can be expanded toward other outputs such as residual stress, fatigue-relevant metrics, or post-processing scenarios. The study demonstrates that a validated physics-based surrogate is a realistic and valuable asset for more efficient AM process development. Acknowledgements The authors acknowledge the European Commission for funding the Twin4Twin project [GA No. 101079180] through the Horizon Europe Framework Programme, Widening Participation and Spreading Excellence initiative.