Laser Metal Deposition with Powder (LMDp)

• LMDp is a powder-fed directed energy deposition process that builds 3D metal structures layer-by-layer by melting metal powders with a focused laser.
• The process relies on precise control of laser power, scan speed, and powder feed to manage melt pool dynamics and ensure accurate deposition geometry.
• Advanced monitoring and modeling, incorporating in-situ sensors and physics-based simulations, enable adaptive defect remediation and improved process efficiency.

Searching arXiv for recent and foundational papers on Laser Metal Deposition with powder (LMDp). Laser Metal Deposition with powder (LMDp) is a directed energy deposition process in which a focused laser beam melts metal powder delivered into the beam–substrate interaction zone, forming a melt pool that solidifies into a track; by scanning this interaction zone and stacking tracks layer-by-layer, three-dimensional structures are built (Donadello et al., 2018). In the literature considered here, LMDp is also treated as the canonical powder-fed configuration of laser directed energy deposition, with coaxial or near-coaxial powder delivery, explicit sensitivity to standoff distance, and strong coupling among laser power, scan speed, powder feed, melt pool dynamics, and deposition geometry (Yan et al., 2022). The process is used for adding features to existing parts, repair, walls and tubes, and multi-axis freeform deposition, and it has become a focal point for in-situ monitoring, multiphysics simulation, inverse optimization, and adaptive quality enhancement (Chen et al., 2024).

1. Process configuration and relation to other laser additive manufacturing routes

In LMDp, a high-power laser is focused onto a substrate through a deposition head while metal powder is fed by a powder feeder, carried by an inert gas, and delivered through a multi-jet coaxial powder nozzle that forms a convergent powder cone. The laser beam and powder cone overlap at the substrate, forming a melt pool; powder particles that enter this melt pool are melted and solidify behind the moving interaction zone to create a solid track. By moving the deposition head over the substrate and incrementally raising the head after each turn or layer, the process builds three-dimensional structures layer-by-layer (Donadello et al., 2018).

Representative systems in the cited works include a 3 kW IPG YLS-3000 fiber laser at 1070 nm with AISI 316L powder of 45–90 µm, and a 3-axis L-DED system with a 1 kW fiber laser at 1070 nm, 1 mm beam diameter, coaxial powder nozzle, argon shielding, and powder feed rate expressed in g/min or as a normalized disc rotation speed with

$$\text{feedrate} = (\text{disc rotation speed}) \times 3.5.$$

Typical variables are laser power $P$, scan speed $v$, powder feed rate $m$, hatch spacing $hs$, layer thickness $t$, and standoff distance between nozzle and workpiece (Shang et al., 2024).

LMDp is distinct from powder bed fusion in both feedstock delivery and process geometry. LMDp uses blown powder dynamically delivered to the melt pool via a nozzle, whereas powder bed fusion uses a static powder bed that is spread layer-by-layer and scanned over the bed. LMDp can deposit on arbitrary freeform surfaces, add features to existing parts, repair, and build walls or tubes using robot or CNC motion; powder bed fusion is limited by bed geometry and the recoating mechanism. A central practical difference is that in LMDp the standoff distance directly influences powder-laser overlap and catchment efficiency, so deviations can produce layer-thickness changes, defects, and dimensional inaccuracies (Donadello et al., 2018).

2. Governing physics: free surfaces, powder catchment, and bead geometry

The process physics of LMDp are commonly formulated as a coupled thermal multi-phase flow problem over gas and metal domains separated by a gas–metal free surface. In one unified formulation, a level-set field $\phi(t,\mathbf{x})$ defines the interface, the interface advects with the local velocity $\mathbf{u}$, and material properties are blended between gas, solid metal, and liquid metal through regularized phase indicators. The momentum balance includes surface tension, Marangoni convection, evaporation acceleration, and recoil pressure, while the energy balance includes latent heat, radiative losses, evaporative cooling, and laser heating obtained from ray tracing. Within this framework, Marangoni flow controls bead width, dilution, and undercut, and recoil pressure strongly affects keyhole depth, bead cross-section, and powder capture at high powers (Yan et al., 2022).

A complementary SPH formulation resolves liquid metal and gas as weakly compressible viscous fluids and treats powder particles as fully resolved rigid bodies that can melt, become part of the fluid melt pool, and later solidify. In that framework, the DED/LMDp example shows melt pool formation in the substrate, powder particles entering the laser interaction zone, partial melting in flight, conduction-driven melting after immersion, deflection of partially molten powder due to recoil pressure, and inclusion of powder when the laser is switched off too early. This supports the interpretation of LMDp as a process in which melt pool thermo-capillarity, evaporation, wetting, and discrete powder dynamics are all first-order phenomena rather than perturbations (Fuchs et al., 2022).

Bead and track geometry have also been formulated through a physics-based, non-empirical deposit geometry model. There, the deposited cross-section $\tilde{\psi}(y)$ is obtained by minimizing surface energy plus gravitational potential subject to mass conservation,

$$\rho_m V_s \int_{-L/2}^{L/2}\tilde{\psi}(y)\,dy = \eta_c \dot{m},$$

with $P$0 the scan speed, $P$1 the catchment efficiency, and $P$2 the powder feed rate. This makes explicit that the deposited area is controlled simultaneously by process kinematics and the mass actually captured by the melt pool (Yan et al., 2022).

Catchment efficiency is therefore a central LMDp variable. In one analytical treatment,

$P$3

and a geometric estimate derived from layer height is

$P$4

A semi-empirical model factorizes the theoretical efficiency into an energetic melting coefficient and a spatial interaction coefficient,

$P$5

linking deposition performance to both available laser energy and laser–powder overlap (Donadello et al., 2021). This suggests that LMDp stability cannot be reduced to laser energy density alone; it is fundamentally mediated by the relative position of melt pool, powder cone, and laser beam.

3. Deposition height, standoff distance, and self-stabilized growth

Deposition height is a process variable of immediate geometric significance because mismatch between programmed robot motion and actual growth produces defects and dimensional errors. In coaxial monitoring studies, the measured relative height is the difference between robot height $P$6 and physical deposition height $P$7,

$P$8

and, at the layer level,

$P$9

where $v$0 is the nominal layer increment. Applied to a stainless steel cylinder built on a helicoidal path with $v$1, this framework revealed a negative transient in $v$2, meaning the structure initially grew faster than the robot height, followed by convergence toward a stationary regime (Donadello et al., 2018).

A related formulation expresses layer mismatch as

$v$3

so that $v$4 implies the part grows faster than the programmed increment and the standoff distance decreases, whereas $v$5 implies the opposite. The cited experiments show that, in certain conditions, LMDp exhibits a passive self-stabilizing regime: initial layers can grow too fast, reducing standoff distance, but that reduction can itself lower powder catchment efficiency by degrading the overlap between the convergent powder cone and the laser-heated region. The process then converges to a standoff distance $v$6 such that the theoretical layer height satisfies

$v$7

In this interpretation, self-stabilization is associated to a reduction of the powder catchment efficiency, governed by the melt pool relative position with respect to the powder cone and the laser beam (Donadello et al., 2021).

This mechanism is not equivalent to exact geometric fidelity from the first layer. The cited studies explicitly show an early transient with overbuilding, and one experiment reports initial layers with $v$8 up to approximately $v$9 larger than the nominal increment before gradual convergence. The practical implication is that self-regulation can maintain a regular height growth and a constant standoff distance after the transient, but it does not eliminate initial dimensional mismatch (Donadello et al., 2018).

4. Monitoring architectures: geometry, melt pool state, and powder quality

A major line of LMDp research concerns in-situ metrology. One direct approach is coaxial laser triangulation integrated into the deposition head. In the reported system, a 532 nm probe beam passes through the nozzle and hits the melt pool, a coaxial camera acquires the probe spot, and spot position is converted to relative height through

$m$0

with calibrated coefficients $m$1 and $m$2. The configuration is coaxial and non-intrusive, does not depend on motion direction, and was demonstrated for reconstructing a spatial map of height variation over a stainless steel cylinder, including stripe-like defects aligned along the helix (Donadello et al., 2018).

The broader monitoring literature for powder-based LDED/LMDp includes coaxial and off-axis visible or NIR cameras, SWIR and MWIR thermal cameras, dual-wavelength pyrometers, microphones, laser line scanners, and operando X-ray monitoring. Physics-informed melt pool features include melt pool area $m$3, ellipse-fit width and length, convex hull area, bounding-box width and length, centroid position, peak temperature $m$4, and moments of the temperature distribution. The review emphasizes that no single sensor captures all relevant defect mechanisms in LMDp, because keyhole pores, lack-of-fusion, cracks, over-build, dents, and powder stream faults have different optical, thermal, acoustic, and geometric manifestations (Chen et al., 2024).

Operando X-ray monitoring addresses the subsurface observability gap. In one industrial LMDp setup, a movable polychromatic X-ray source and flat-panel detector were used to image ten successive layers of SS316L. The physical basis is the density difference between liquid and solid phases, modeled for SS316L as

$m$5

combined with Beer–Lambert attenuation,

$m$6

Traditional image analysis failed because the polychromatic beam rendered contrast variations difficult to detect, but a VGG16-inspired auto-encoder trained on simulated thermal and X-ray data showed promising localization of the melt pool in low-contrast radioscopic images (Jegou et al., 26 Sep 2025).

Monitoring in LMDp also extends upstream to feedstock characterization. Laser induced breakdown spectroscopy has been used for quantitative multielement analysis of additive-manufacturing powders, including heterogeneous WC–Ni mixtures and light elements such as carbon. For tungsten, the calibration curve

$m$7

yielded $m$8 and $m$9, while carbon quantification using the C I 193.09 nm line gave $hs$0 and $hs$1. This provides a route for express on-site multielement analysis of powders used in LMDp (Lednev et al., 2018).

5. Modeling, inverse design, and fast predictive surrogates

The current modeling landscape for LMDp spans high-fidelity multiphase simulation, mesoscale particle methods, data-driven inverse optimization, and physics-informed deep learning. Thermal multi-phase flow solvers with mixed interface-capturing and interface-tracking reconstruct the gas–metal free surface through explicit triangulation, enable ray tracing with multiple reflections, and predict quantities that experiments cannot measure directly, including melt pool flow patterns, centerline thermal fields, cooling rates, and proxies such as secondary dendrite arm spacing and hardness. For SS-316L single-track DED, one such model reported quasi-steady melt pool dimensions $hs$2, $hs$3, $hs$4, and hardness $hs$5 (Yan et al., 2022).

SPH provides a different route in which laser–powder–melt interactions are resolved with discrete powder particles, free-surface deformation, wetting, and recoil pressure. That framework is especially suitable for mesoscale modeling of dynamically changing interface topologies, and it directly captures powder motion, distortion of powder packing structure, powder particle ejection, and inclusion of partially molten particles under unfavorable process timing (Fuchs et al., 2022).

Inverse process optimization has moved toward contour-based machine-learning surrogates. The AIDED framework combines forward MLP models, geometric reconstruction, and the NSGA-III algorithm in Pymoo to predict cross-sectional melt pool geometry from process parameters and then identify process parameters for specified objectives. Reported performance includes $hs$6 for single-track melt pool area prediction on 316L, $hs$7 for multi-track tilt angle prediction, and multi-layer prediction errors of $hs$8 and $hs$9 in width and height, respectively. The same framework was used to determine hatch spacing and layer thickness for fully dense prints with density $t$0, and to inversely identify optimal process parameters within $t$1–$t$2 hours; transfer learning from stainless steel to pure nickel improved single-track area prediction on Ni from $t$3 to $t$4 (Shang et al., 2024).

A different surrogate strategy is thermo-mechanical physics-informed deep learning. In a Ti-6Al-4V single-layer LMD study, a thermoelastic PINN encoded the transient heat equation and linear thermoelastic equilibrium directly into the network loss. The temperature field at $t$5 matched FEM with RMSE $t$6; a full FEM run took approximately $t$7 hours on a single CPU, whereas transfer of a pre-trained PINN to new laser power and scan speed converged in approximately $t$8 minutes. The authors explicitly frame this as online soft sensing for fast prediction of temperature and thermal stress evolution during LMD (Sharma et al., 2024).

6. Defects, quality assurance, and adaptive process control

The defect landscape in LMDp spans gas-induced pores, keyhole pores, lack-of-fusion pores, cracks, microstructure inhomogeneity, layer unevenness, over-build, under-build, warpage, distortions, and surface roughness. The review emphasizes that some of these mechanisms are specific to powder-based LDED, notably gas entrapment from powder and shielding gas, and geometric fluctuations caused by changing nozzle–melt pool distance (Chen et al., 2024).

Machine-learning-assisted defect detection has been demonstrated across several modalities. In LMDp acoustics, MFCC-CNN pipelines have been used for classification of defect-free deposition, cracks, and keyhole pores; one study reported overall accuracy of approximately $t$9, keyhole pore detection accuracy of approximately $\phi(t,\mathbf{x})$0, and AUC-ROC of approximately $\phi(t,\mathbf{x})$1. In multisensor robotic LMDp, feature-level fusion of visual, thermal, and acoustic streams improved defect classification performance and reduced false alarm rates. These results support a shift from single-sensor thresholding toward synchronized, robot-registered quality maps (Chen et al., 2024).

Control strategies in the cited literature operate on melt pool width, melt pool temperature, or deposition height through laser power, scan speed, and powder feed rate. A generic controller is written as

$\phi(t,\mathbf{x})$2

with $\phi(t,\mathbf{x})$3 defined from a target melt pool width or temperature. The same review also describes model predictive control for height evolution, layer-height compensation, repetitive control, and hybrid additive–subtractive remediation in which laser line scanning identifies bulges and dents and corrective toolpaths are generated accordingly (Chen et al., 2024).

A persistent misconception is that any apparent self-stabilization of LMDp necessarily implies optimal process efficiency. The cited evidence shows the opposite: stabilization can be accompanied by a reduction in powder catchment efficiency, and the early transient can still generate dimensional mismatch or overbuilding (Donadello et al., 2021). Another practical implication is that direct geometric sensing and indirect thermal or acoustic sensing are complementary rather than interchangeable: triangulation measures height directly, whereas pyrometry, imaging, and acoustic methods often require model-based or learned correlations to geometry or defect state (Donadello et al., 2018).

Across the cited studies, the trajectory for LMDp is clear. Monitoring systems are moving from isolated sensors toward multimodal fusion; models are moving from single-track calibration toward predictive multiphysics and transferable surrogates; and control is moving from open-loop parameter selection toward adaptive defect remediation. This suggests a convergence toward LMDp platforms in which standoff distance, melt pool geometry, powder quality, and part-scale thermal history are all treated as measurable or inferable state variables rather than latent disturbances (Jegou et al., 26 Sep 2025).