Multiscale characterization and controllability by laser-ultrasounds of WLAM components: toward a physics-based and machine learning enhanced online monitoring – COLUMBO

The COLUMBO project adopted an integrated framework combining multiphysics modeling, advanced experimental characterization, and physics-informed machine learning to establish quantitative relationships between process parameters, melt-pool dynamics, microstructure, and ultrasonic response.

The WLAM process was modeled using coupled thermo-hydraulic simulations that predict melt-pool evolution, wire melting, and bead formation. The resulting thermal histories are coupled to cellular-automaton–finite-element (CA-FE) models to predict grain growth, crystallographic texture, and elastic anisotropy. Ultrasonic wave propagation in the resulting heterogeneous media was described through complementary analytical and numerical approaches. Multiple-scattering theory based on the second-order Dyson equation captures attenuation and dispersion in statistically heterogeneous media, while discontinuous Galerkin simulations on realistic microstructures resolve bulk- and surface-wave propagation, including anisotropic and directional effects.

These models were validated experimentally using an instrumented WLAM platform combining high-speed imaging, Schlieren diagnostics, particle image velocimetry, laser interferometry, and laser ultrasonics. A robotic inspection head integrating coaxial laser generation and detection was developed to enables automated in situ measurements during fabrication.

Ultrasonic signals were analyzed using SAFT laser imaging with surface and bulk waves, spatiotemporal Fourier filtering, and time-frequency analysis, and others to extract physically relevant observables, including phase velocity, attenuation, and wave directivity was the subject of the study.

To enable real-time inversion, physics-informed surrogate models based on Fourier Neural Operators and convolutional encoder-decoder architectures accelerated forward simulations by more than two orders of magnitude while reconstructing microstructural descriptors directly from ultrasonic measurements.

These developments have provided the computational foundation for digital-twin-assisted in situ monitoring and real-time quality control of WLAM.

COLUMBO project has established an integrated framework linking the WLAM process, microstructure evolution, and laser-ultrasonic response, providing the scientific basis for in situ monitoring and digital-twin-assisted quality control.

Multiphysics thermo-hydraulic models based on a finite-element volumetric heat-source formulation were developed to simulate melt-pool evolution and bead formation. The proposed approach reduced computational time from several weeks to a few days while preserving predictive accuracy. Coupling these simulations with cellular-automaton–finite-element (CA-FE) models enabled quantitative prediction of grain growth, crystallographic texture, and elastic anisotropy from local thermal histories, thereby establishing direct relationships between processing conditions and the resulting microstructure.

Ultrasonic wave propagation in the resulting heterogeneous media was investigated through complementary analytical and numerical approaches. Multiple-scattering theory based on the second-order Dyson equation, together with high-order discontinuous Galerkin simulations on realistic microstructures, accurately reproduced bulk- and Rayleigh-wave propagation, revealing the strong directional dependence of attenuation and dispersion induced by the anisotropic microstructure.

An advanced experimental WLAM platform was developed by integrating high-speed imaging, Schlieren diagnostics, particle image velocimetry, laser interferometry, and laser ultrasonics. A compact robotic optical head enabling coaxial laser generation and detection in both point- and line-source configurations provided quantitative measurements of melt-pool dynamics and ultrasonic responses under realistic processing conditions.

To enable real-time monitoring, physics-informed surrogate models combining Fourier Neural Operators and convolutional encoder-decoder networks were developed, accelerating ultrasonic simulations by more than two orders of magnitude while maintaining excellent predictive capability. Initial inverse-modeling results demonstrate the feasibility of reconstructing microstructural descriptors directly from ultrasonic measurements.

Toward a digital twin. A 3D model of the experimental cell, a closed-loop architecture integrating stereo vision, laser-plane sensing, and the IRC5 controller, together with a NAS/Kafka data infrastructure, have been established, providing the foundation for in situ LU experiments. Collectively, these advances establish a quantitative framework linking process parameters, microstructure evolution, and ultrasonic signatures, laying the foundation for digital-twin-assisted in situ monitoring and nondestructive characterization of wire laser additive manufacturing (WLAM) components.

The results obtained within COLUMBO open several key scientific and technological perspectives toward robust in situ monitoring of WLAM and the development of a digital twin integrating process modelling, microstructure prediction, and nondestructive evaluation.

Refinement of multiphysics models in multi-bead and multi-layer configurations is a primary objective, requiring improved treatment of cumulative thermal effects, transient phenomena, and preheated wire kinematics (H-WLAM) to enhance predictive capability under realistic processing conditions.

On the materials side, extending microstructural modeling to fully three-dimensional heterogeneous geometries is essential. This effort will rely on the integration of advanced experimental characterization techniques, including EBSD, with high-fidelity process simulations, in order to strengthen the quantitative link between solidification conditions, crystallographic texture, and elastic anisotropy.

For ultrasonic nondestructive evaluation, extending both analytical and numerical frameworks to full 3D configurations, including coupled bulk and Rayleigh wave propagation, remains a central challenge. In particular, the separation of coherent and incoherent contributions within strongly scattering signals is a key bottleneck for reliable interpretation in additively manufactured materials.

In parallel, inverse problem strategies must be made more robust with respect to experimental variability. Embedding physical constraints within machine learning architectures through physics-informed approaches offers a promising route to improve generalization, interpretability, and transferability between simulated and experimental datasets.

Experimental integration is expected to evolve toward fully automated platforms combining 3D high-speed imaging, robot-mounted laser-ultrasonic heads, and GPU-accelerated real-time processing pipelines, enabling a step change in measurement throughput and system autonomy.

Digital twin. A 3D model of the experimental cell, a closed-loop architecture integrating stereo vision, laser-plane sensing, and the IRC5 controller, together with a NAS/Kafka data infrastructure, have been established as the first concrete building blocks toward a digital twin of the WLAM process. Ultimately, these developments converge into a unified framework coupling multiphysics modeling, microstructure evolution, ultrasonic wave propagation, and machine learning. This integrated capability will support real-time monitoring and decision-making strategies for metal additive manufacturing, with strong application potential in aerospace, energy, and advanced industrial sectors.

The industrial use of components manufactured using additive manufacturing (AM) processes has strong growth potential in various challenging fields such as aeronautics, automotive, medical or nuclear. The desire to produce structural parts implies that they are inevitably subjected to Non-Destructive Testing (NDT) and that it is essential to master the process during manufacturing to give the parts a microstructure with the characteristics that make them usable. Implementing an online NDT monitoring strategy would lead to more efficient and optimal control by acting on the AM parameters to avoid process drifts or by repelling defective parts as soon as possible. However, despite increasingly developed studies on the influence of AM parameters on the resulting microstructures, actual knowledge remains incomplete, particularly for the wire-laser process (WLAM), which is more recent and of which certain advantages arouse growing interest among industrials. The availability of an online NDT would allow a significant advance in AM.

COLUMBO aims to demonstrate the ultrasonic-laser (UL) controllability of WLAM parts by quantifying the material parameters that make these controls possible and by defining the detectable characteristic quantities, in order to prove the feasibility of an effective online monitoring strategy. Ultrasonic NDT methods, proven by their fundamental sensitivity to local mechanical characteristics, are methods of choice both for the flaws probing (heat-affected zones, microporosities or cracks) and for the multiscale characterization of microstructures. Besides, UL techniques allow contactless inspection in hostile environments, such as the AM. Finally, the use of Rayleigh waves seems well suited to successively control and characterize the deposited metal layer-by-layer. It is clear that the efficiency of such an NDT procedure can only be guaranteed subject to a relevant physical interpretation and optimal use of the data of the real-time online control data. However, the WLAM parts constitute a challenge because of the complex phenomena of ultrasonic diffusion linked to their very marked microstructures (surface roughness, porosities, entanglement of columnar/equiaxial grains), very different from those resulting from conventional metallurgical processes. The associated signals, potentially rich in information, are therefore complex to interpret. The scientific issue to be addressed is the mastery of the correlation between the WLAM parameters, the characteristics of the obtained microstructure and its ultrasonic signature.

COLUMBO proposes to meet this challenge by developing multiscale and multi-physics modelling/ simulations of the WLAM process and the ultrasound propagation, both by comparing with characterization data and experimental measurements. Thanks to the manufacturing/measurement/modelling complementarity between the five partners, the considered methodology consists of establishing a hybrid benchmark of the carefully chosen parts with a microstructure of increasing complexity and with sets of well-identified and classified WLAM parameters, to continuously advance the knowledge on involved phenomena, the modelling of underlying physical mechanisms, and the quantification of measurable quantities. The ultimate goal is an optimal exploitation of real-time in situ control data using simplified models with validated physical content, and based on machine learning (ML).

To achieve this ambitious and topical objective, COLUMBO brings together a consortium of five partners with skills that fulfil the entire chain of expertise required: from WLAM process modelling/simulation, to online testing of WLAM process, including ultrasound modelling/simulation and characterization (EBSD), towards inversion of data by ML. This complementarity with a fair balance between theoretical/numerical modelling and experimental validation constitutes the key point of the rigor and the success of COLUMBO.