Wire-Laser Additive Manufacturing (WLAM)

• WLAM is a wire-fed laser deposition process that builds parts layer by layer while enabling high automation, rapid deposition, and integrated sensing for quality control.
• Process parameters such as laser power, travel speed, and wire feed rate dynamically influence molten-pool characteristics, bead geometry, and thermal histories.
• Advanced control strategies, including in situ sensing, CNN-based inference, and digital-twin formulations, support real-time process optimization and defect management.

Wire-Laser Additive Manufacturing (WLAM), also described in the literature as wire-feed laser additive manufacturing and Laser Directed Energy Deposition of Wire (DED-LB/w), is a wire-based directed energy deposition process in which a laser melts incoming wire feedstock onto a substrate to build parts layer by layer. The process is associated with high automation, high deposition rates, and suitability for large-scale production, but its practical deployment is constrained by strongly coupled process variables, evolving molten-pool dynamics, residual-stress generation, and defect-sensitive thermal histories. Across recent work, WLAM is therefore treated not merely as a deposition method, but as a coupled thermo-fluid-mechanical and data-centric manufacturing system in which sensing, inference, and control are integral to part quality (Jamnikar et al., 2021, Liu et al., 2021, Kannapinn et al., 2024).

1. Nomenclature, scope, and relation to neighboring processes

In the cited literature, WLAM denotes a laser-heated, wire-fed additive process within the broader DED family. One formulation studies a laser hot-wire DED/WLAM system for Ti-6Al-4V; another considers Laser Directed Energy Deposition of Wire (DED-LB/w) for digital-twin development on Inconel 718. Despite differences in application focus, both treat the process as laser-based wire deposition with layerwise buildup, strong thermal transients, and process-state observability through field sensing or simulation (Jamnikar et al., 2021, Kannapinn et al., 2024).

A recurring terminological issue is conflation of WLAM with Wire + Arc Additive Manufacturing (WAAM). The tungsten study "Development of Wire + Arc Additive Manufacturing for the production of large-scale unalloyed tungsten components" is not laser-based WLAM: it uses a conventional TIG torch rather than a laser heat source. That distinction matters because WLAM and WAAM differ in heat-source physics, plume behavior, and transfer conditions. At the same time, the WAAM tungsten work is directly informative for wire-fed additive manufacturing more generally because it addresses feedstock transfer stability, heat-input control, defect suppression, layer geometry control, and scaling to large refractory-metal parts (Marinelli et al., 2019).

WLAM must also be distinguished from powder-based laser AM. Several tungsten studies in the broader laser AM literature are conducted in laser powder bed fusion rather than wire-fed deposition. Their relevance is nevertheless substantial, because cracking, porosity, residual-stress buildup, recrystallization, and alloy-design strategies transfer at the level of materials physics even when the feedstock delivery mode differs (Cunningham et al., 2023).

2. Process variables, molten-pool state, and bead formation

A central theme in WLAM research is that the molten pool is the most informative local representation of process state. In Ti-6Al-4V WLAM, molten-pool dimensional information and temperature are treated as direct indicators of bead quality, thermal history, microstructure, residual stress, and defects. The principal controllable variables studied are laser power (LP), travel speed (TS) or robot travel speed (RTS), wire feed rate (WFR) or wire feed speed (WFS), and hot wire power (HWP) or wire heat power (WHP); oxygen content and stage-specific machine settings also appear in broader quality studies (Jamnikar et al., 2021, Liu et al., 2021).

The experimental parameter study on single-bead WLAM deposits at Oak Ridge National Laboratory reports LP in the range 4000–6000 W, TS in the range 3.5–10 mm/s, WFR in the range 40–71.3 mm/s, and nominal HWP of 300 W. The measured responses show that increasing LP from 4000 to 5000 W enlarges the molten pool and raises molten-pool temperature from about 1800°C to 1850°C to 1925°C. By contrast, increasing TS slightly decreases molten-pool length and has a weaker temperature effect, while increasing WFR tends to reduce molten-pool temperature because the incoming solid wire contributes a cooling effect. Step-change experiments further indicate a control-action latency: a $\pm 500$ W laser-power change produces about $\pm 50^\circ$C change in molten-pool temperature, and the molten pool typically reaches a new steady state in about 3–5 s (Jamnikar et al., 2021).

The broader quality study of 179 single-layer WLAM beads formalizes bead description through geometric and derived metrics. The primary cross-sectional quantities are bead height $h$, bead width $w$, fusion zone depth $d$, and fusion zone area $A$. Two derived descriptors are

$Du=\frac{d}{h+d}, \qquad AR=\frac{w}{h}.$

Within the learned process space, increasing RTS decreases $h$, $w$, $d$, and $\pm 50^\circ$0; increasing LP generally increases bead width, fusion zone depth, and fusion zone area, but can decrease bead height because hotter material spreads more; increasing WFS generally increases bead height and often fusion zone area, with a strongly interaction-dependent effect on width (Liu et al., 2021).

These results establish the physical basis for later monitoring and control work. They indicate that WLAM quality is not governed by any single scalar heat-input descriptor, but by interacting energy, feed, and motion variables whose effects are mediated by molten-pool morphology and thermal state.

3. In situ sensing and learning-based inference

The sensorized WLAM platform reported in the Ti-6Al-4V studies uses a 6 kW laser, 1.5875 mm Ti-6Al-4V wire feedstock, and an argon-filled environment. Two Prosilica GT1930C cameras are mounted on the robot head, one coaxial/in-line and one oblique at $\pm 50^\circ$1, and three pyrometers observe molten-pool, leading-edge, and trailing-edge temperatures. The image stream is recorded at 25 fps with raw resolution $\pm 50^\circ$2 and cropped to a region of interest of $\pm 50^\circ$3 for learning; the pyrometers are sampled at 100 Hz. In the process-parameter inference study, the molten-pool pyrometer is treated as the most informative thermal feature, outperforming leading-edge and trailing-edge temperatures in regression performance (Jamnikar et al., 2021).

That work defines a sensing-process (S-P) model: given molten-pool image data, optionally augmented with molten-pool temperature, predict the underlying process parameters $\pm 50^\circ$4. The uni-modality CNN uses image input only and has 31 layers, 7 convolution layers, 3 global average pooling layers, batch normalization, tanh activation, and 30% dropout. The multi-modality CNN performs late fusion of image features with molten-pool temperature and has 24 layers total, 4 convolution layers, 2 global average pooling layers, batch normalization, tanh activation, and 50% dropout. On a dataset of 6500 images, with 6000 training samples and 500 unseen test samples, the multi-modality model improves RMSE from 385.23 to 186.48 for LP, from 1.54 to 1.12 for TS, and from 7.97 to 6.41 for WFR, with reported relative errors of 3.49% for LP, 26.70% for TS, and 13.93% for WFR (Jamnikar et al., 2021).

A complementary study defines a sensing-quality (S-Q) model for in situ quality estimation rather than parameter inference. Here the same sensor classes—molten-pool images and pyrometer temperature data—are mapped directly to post-process bead geometry and microstructural quantities. The geometric outputs are bead height $\pm 50^\circ$5, bead width $\pm 50^\circ$6, fusion zone depth $\pm 50^\circ$7, and fusion zone area $\pm 50^\circ$8; the microstructural outputs are alpha lath thickness $\pm 50^\circ$9, beta grain length parallel to the build $h$0, and beta grain length perpendicular to the build $h$1. The CNN is again multi-modal, with separate geometry and microstructure models built around a 24-layer architecture. On 13 single-bead experiments, using 12 builds for training and 1 unseen build for testing, the geometric CNN reports RMSE values of 0.30 mm for $h$2, 0.50 mm for $h$3, 0.09 mm for $h$4, and 3.39 mm$h$5 for $h$6; the microstructural CNN reports RMSE values of 0.02 $h$7m for $h$8, 20.27 $h$9m for $w$0, and 30.26 $w$1m for $w$2. The reported interpretation is that molten-pool image morphology and temperature jointly encode sufficient state information to estimate both geometric and microstructural outcomes during deposition (Jamnikar et al., 2021).

Taken together, these studies define a layered inference stack for WLAM: process parameters $w$3 molten-pool state $w$4 geometry and microstructure. They also show that temperature is not a minor auxiliary feature; it provides information absent from image morphology alone.

4. Closed-loop control and digital-twin formulations

The move from in situ estimation to active control appears in two related strands: layerwise geometric correction and faster-than-real-time field prediction. In multi-robot WAAM, a three-robot “scan-n-print” framework integrates deposition, scanning, reconstruction, and next-layer correction. The system consists of a 6-DOF welding robot, a 2-DOF trunnion positioner, and a 6-DOF sensing robot carrying a wrist-mounted laser line scanner. Welding parameters, including wire feed rate, are held constant, so the control input is robot path speed; the measured output is the part height profile. The identified deposition model is

$w$5

and the next-layer segment speed is obtained by model inversion,

$w$6

On a flat wall, the closed-loop method improves height standard deviation by 66% over all layers, by 60% when edge regions are excluded, average CAD distance error by 13%, and maximum error by 72%. On a turbine-blade-like geometry, height standard deviation improves by 53%, tracking RMSE by 45%, average CAD distance error by 13%, and maximum error by 5% (Lu et al., 2024).

Although that work is arc-based rather than laser-based, its control structure is directly transferable at the workflow level: print, scan, reconstruct, compare to target, and adjust the next deposition command. This suggests that WLAM can adopt comparable geometry-control architectures even when the actuation variables differ, for example through coordinated adjustment of laser power, scan speed, wire feed rate, or tool-path strategy.

A more explicitly WLAM-centered control precursor is the DED-LB/w digital-twin study based on multi-physical simulation and neural ordinary differential equations. The physical model simulates a single laser line scan over an Inconel 718 block of dimensions $w$7 cm using transient heat conduction and quasi-static small-strain thermo-elasto-plasticity. The learned surrogate is split into two stages, $w$8 for temperature fields and $w$9 for stress fields, with latent dynamics governed by

$d$0

The finite-element model has about 2.26 million degrees of freedom; a 20 s real-time process takes about 10 minutes on 16 CPU cores. The surrogate predicts the full 20 s temperature field in about 412 ms and the stress field in about 410 ms on a standard desktop CPU, with relative errors consistently below 1% for both temperature and stress over time and best performance at latent dimension $d$1 (Kannapinn et al., 2024).

The digital-twin formulation is still a proof of concept: it does not yet include actual material deposition and is restricted to a single line scan with fixed process parameters. Even so, it establishes a core proposition for WLAM control research: field-level prediction can be faster than the physical process, and thus can support on-the-fly re-optimization rather than purely retrospective correction.

5. Defects, geometric metrology, and image-based inspection

Defect management in WLAM includes both volumetric/internal anomalies and external geometric deviation. A metrology-focused WLAM study addresses the latter for parts composed of multiple entities, emphasizing that an early geometric error can propagate into positioning, machining allowance, and subsequent entity production. The proposed solution combines stereoscopic imaging with a superellipsoid-based parametric model and global image stereocorrelation. The superellipsoid surface is defined as

$d$2

with three scale parameters $d$3 and two shape parameters $d$4. In the reported case study, the nominal part is a $d$5 mm $d$6 $d$7 mm $d$8 $d$9 mm cuboid with rounded edges of radius 15 mm, corresponding nominally to $A$0, $A$1, and $A$2. After in-situ optimization, the recovered parameters are $A$3; convergence occurs in nine iterations, and comparison with an ATOS Core scan gives a standard deviation of 0.34 mm. The stated significance is not capture of every local surface irregularity, but recovery of a compact parameterized surface that can update downstream CAM operations (Hachem et al., 22 Sep 2025).

A different inspection problem is addressed by the promptable defect-segmentation study built around the Segment Anything Model (SAM). The demonstrated case is XCT-based porosity segmentation in laser powder bed fusion rather than WLAM, but the method is presented as a label-free, low-latency defect-segmentation workflow for laser additive manufacturing more broadly. Images are clustered with K-means, centroid images are thresholded into foreground and background, and a random subset of foreground coordinates is supplied to SAM as multi-point prompts. Without prompts, SAM yields ambiguous defect masks with Dice Similarity Coefficient values between 0.07 and 0.48. With centroid-based prompts, average DSC reaches 0.81 on one image set from specimen A and 0.75 on another; increasing prompt percentage from 0.1% to 18% improves DSC from 0.697 to 0.829, with 18% limited by memory constraints. Runtime for a single RGB image of size $A$4 with multiple point prompts is reported as 3.8 to 4.4 seconds (Era et al., 2023).

For WLAM, the segmentation paper is relevant as a transferable inspection framework rather than a validated in-process solution. The authors explicitly note that adaptation would be required for different image domains, defect morphologies, and prompt-selection policies. This directly counters a common misconception that foundation-model-based defect segmentation is plug-and-play across laser AM modalities.

6. Refractory materials, tungsten, and wire-fed process lessons

Tungsten concentrates several of the most severe materials challenges faced by laser-based additive manufacturing. The laser AM tungsten study reports solidification cracking or hot cracking, embrittling grain-boundary features such as oxides and pores, large thermal expansion strains during rapid laser heating and cooling, repeated thermal cycling through the ductile-to-brittle transition temperature of about 723 K, and recrystallization at about 1100–1300°C. To address these issues, four L-PBF compositions are compared: pure W, W + 0.5 ZrC, W-3.5Ni-1.5Fe, and W-3.5Ni-1.5Fe + 0.5 ZrC. ZrC promotes grain refinement and microstructural stability through ZrO$A$5/ZrO dispersoids, while NiFe forms a ductile FCC NiFeW phase within the BCC W matrix. The reported hardness values are 420 HV for pure W, 460 HV for W + ZrC, 530 HV for W-3.5Ni-1.5Fe, and 595 HV for W-3.5Ni-1.5Fe + 0.5 ZrC. Pure W and W + ZrC show indentation microcracks, whereas WNiFe and WNiFe + ZrC do not. The combined alloy is presented as balancing crack suppression, hardness, and thermal stability more effectively than either binary alloy alone (Cunningham et al., 2023).

That powder-based laser study does not use wire feedstock, but its implications for WLAM are direct at the materials-design level. It shows that tungsten printability in laser AM cannot be treated as a purely process-parameter problem; phase constitution, grain-boundary chemistry, and dispersoid-mediated stabilization are equally important control variables.

The neighboring wire-fed evidence is provided by the TIG-based tungsten WAAM study. Using unalloyed tungsten wire of 1 mm diameter, an airtight enclosure with oxygen reduced to about 100 ppm, 100% helium shielding gas, and a clamping force of 25 N, the study shows that substrate orientation and wire-feeding orientation decisively affect defect formation. Flat-position deposition caused consistent cracking, whereas deposition on the edge of the substrate prevented lateral cracking. More importantly, side wire feeding produced spattering, voids, lack of fusion, porosity, and micro-cracks, which the authors relate to Kelvin–Helmholtz instability at the interface between liquid tungsten and the high-velocity helium plasma. Front wire feeding produced no visible spatter, stable liquid bridge transfer, no significant weld-pool perturbation, and defect-free deposits in terms of pores and lack of fusion. A relatively large-scale component of $A$6 mm was fabricated, and the authors conclude that Wire + Arc Additive Manufacturing can be a real candidate to replace conventional manufacturing of fully dense large-scale tungsten components (Marinelli et al., 2019).

For WLAM, this neighboring result does not establish laser-specific tungsten process windows, but it does establish a robust wire-fed principle: transfer stability, wire orientation, and thermal management govern whether refractory-metal deposition remains defect-free as scale increases. A plausible implication is that tungsten WLAM must solve the same coupled transfer and thermal-shock problems, even though the governing flow field is laser-driven rather than arc-driven.