PSI - Issue 83

C. Mallor et al. / Procedia Structural Integrity 83 (2026) 130–137

131

1. Introduction Additive manufacturing (AM) has evolved from a niche prototyping capability into an enabling production technology for sectors that require geometric freedom, lightweight structures, and short design-to-part cycle. Among the available metal AM routes, powder bed fusion using a laser beam for metals (PBF-LB/M) [1] is especially relevant because it can create intricate metallic components layer by layer from digital definitions while keeping material use comparatively efficient. Its adoption in aerospace and automotive [2], energy [3], and electronics applications [4] continues to grow because the process can achieve features that are difficult to obtain with conventional subtractive routes. However, there are still several challenges that delay its general adoption [5]. The same attributes that make PBF-LB/M attractive also make it difficult to industrialize robustly. During manufacturing, thermal gradients, repeated heating and cooling, and subsequent stress relaxation can generate residual stress, distortion, and warpage. The interaction of part geometry, support design, scanning strategy, and thermal boundary conditions means that a process window that works for one component may not transfer cleanly to another. In practice, this has often led to expensive trial-and-error campaigns and delayed process qualification, which remain a barrier to wider adoption [6]. Physics-based simulation has become one of the most important tools for process development in metal AM. Finite element models can represent evolving heat input, thermal losses, material deposition, and structural response, making them valuable for understanding how process settings affect the final built shape. In the AM literature, such models range from inherent-strain approaches to detailed thermo-mechanical frameworks that attempt to preserve more of the underlying physics [7–9]. Even when macro-scale abstractions are used, high-fidelity models remain computationally demanding and are not always suitable for fast exploration, real-time estimation, or to be used in decision-support tools. This computational bottleneck has motivated strong interest in digital twins (DTs), surrogate models (SMs), and reduced-order models (ROMs) for AM. A DT needs digital representations that are accurate enough to remain physically meaningful but fast enough to support monitoring, optimization, and decision-making. In that context, surrogate models act as efficient substitutes for the more expensive experimental or simulation layers. ROM strategies are particularly appealing because they do not merely fit outputs; they also retain some structure of the physical problem and can preserve interpretability when the model is built carefully [10–13]. Recent literature shows detailed thermo-mechanical modelling continues to improve for residual stress and distortion prediction [14–16]. Digital-twin frameworks are becoming more mature for process monitoring, control, and integration [17–19]. At the same time, surrogate approaches have been used to accelerate calibration, predict residual stress, or estimate distortion [20–22]. Literature still lacks benchmark studies that explicitly relate the three layers in a single workflow: a well-defined experimental set-up, a physics-based simulation framework, and an interpretable surrogate model trained on validated numerical data. That integrated perspective matters because the credibility of any fast model depends on the quality of the benchmark that generated it. Without a traceable link between the printing conditions, measurements, modelling assumptions, and the ROM architecture, it is difficult to assess how much trust should be placed in a surrogate prediction when it is deployed beyond the initial test case. The objective of this study is to develop and assess a physics-based surrogate model for the distortion of a metal twin-cantilever benchmark manufactured by PBF-LB/M. Three process parameters are considered: laser power, scan speed, and build-plate preheating temperature. The study combines physical builds on an open-architecture machine, in-situ and ex-situ measurements, a sequential thermo-mechanical simulation framework, and a tensor-rank decomposition-based ROM that can be interpreted and eventually integrated within a DT environment. The present article revisits the original study [23] from a simpler perspective, reusing the main datasets[24]. The central question is whether a validated macro-scale physics-based model can be condensed into an efficient surrogate without losing the essential trends that matter for process parameters selection. 2. Physics-based surrogate model methodology This section describes the experimental procedure, simulation strategy framework, and surrogate model (SM) workflow within a physics-based methodology.