Approche hybride IA/CFD pour la simulation des impacts de gouttes lors des procédés de projection thermique // Hybrid AI/Stochastic/CFD Approach for Simulating Droplet Impacts in Thermal Coatings

Suspension Plasma Spraying (SPS) is an emerging industrial process, particularly for the creation of ceramic coatings resistant to thermomechanical stresses, used as long-life thermal barriers for aircraft engine. For the aeronautics industry, it is classified as a special process whose output elements can only be verified by monitoring or post-measurement, and whose deficiencies therefore only become apparent once the product is in use.

In this process, a liquid suspension containing submicron particles of the material to be deposited is injected into a thermal plasma jet, where it is fragmented and evaporated. This releases individual or agglomerated submicron particles, which are subsequently accelerated and melted before impacting the surface of the part to be coated, spreading upon impact and solidifying to build up the coating.

The structure of the coating is a function of the operating conditions, from the plasma torch to the droplet impact conditions (shape, velocity, temperature, and substrate roughness). A dense or columnar structure may occur, which influences the final thermomechanical properties of the material. A full CFD simulation of the entire process is beyond reach due to limitations in the number of particles that can be simulated. Therefore, we propose a multi-step approach, consisting of CFD simulations at the droplet scale combined with a stochastic approach [1] enriched by AI at the coating scale:

• Simulations of droplet impacts using the CFD code Notus [2] aim to populate a database representing the topology of various instantaneous, representative sprayed surfaces, ranging from simple to more complex configurations;
• A neural network tool, trained on CFD results, aims to replace CFD simulation capacity by representing impact surfaces with sufficient accuracy. The AI tool's results can be validated and refined through additional CFD simulations;
• The stochastic approach aims to represent realistic spray conditions by modeling the spatio-temporal distribution of particles, including their radius, velocity, temperature, and impact point;
• The combination of the stochastic approach and the AI tool will surpass the capabilities of CFD simulations by enabling the representation of large surfaces and large numbers of particles.

The PhD thesis focuses on the AI/CFD component of the project. Artificial intelligence has been used for several years to detect, segment and reconstruct the contour of the interface between two fluids in experimental images [3, 4]. Rather than segmentation or object detection in an image, the objective of the AI tool to be developed is the prediction of a final state from an initial state and relevant parameters. Given an initial condition consisting of a field of volume fractions representative of the surface state and a characterisation of the particle and its point of impact before impact, the tool must predict the new volume fraction state representative of the droplet's spreading and subsequent solidification. The work will start with a bibliography of available AI analysis tools based on image and volume fraction processing. It will also involve creating a database of droplets impacting substrates with an increasing level of topological complexity, Python programming, comparing the selected methods and 2D/3D verification of the proposed approaches. Particular attention will be paid to quantifying the reliability of predictions [5].

Please send a detailed CV, undergraduate and Master's transcripts, Master's reports and reference contacts to all the following contacts:

Emmanuelle Abisset-Chavanne I2M-Bordeaux : emmanuelle.abisset-chavanne@ensam.eu
Stéphane Glockner I2M-Bordeaux: glockner@bordeaux-inp.fr
Vincent Rat IRCER-Limoges: vincent.rat@unilim.fr
------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Suspension Plasma Spraying (SPS) is an emerging industrial process, particularly for the creation of ceramic coatings resistant to thermomechanical stresses, used as long-life thermal barriers for aircraft engine. For the aeronautics industry, it is classified as a special process whose output elements can only be verified by monitoring or post-measurement, and whose deficiencies therefore only become apparent once the product is in use.

In this process, a liquid suspension containing submicron particles of the material to be deposited is injected into a thermal plasma jet, where it is fragmented and evaporated. This releases individual or agglomerated submicron particles, which are subsequently accelerated and melted before impacting the surface of the part to be coated, spreading upon impact and solidifying to build up the coating.

The structure of the coating is a function of the operating conditions, from the plasma torch to the droplet impact conditions (shape, velocity, temperature, and substrate roughness). A dense or columnar structure may occur, which influences the final thermomechanical properties of the material. A full CFD simulation of the entire process is beyond reach due to limitations in the number of particles that can be simulated. Therefore, we propose a multi-step approach, consisting of CFD simulations at the droplet scale combined with a stochastic approach [1] enriched by AI at the coating scale:

• Simulations of droplet impacts using the CFD code Notus [2] aim to populate a database representing the topology of various instantaneous, representative sprayed surfaces, ranging from simple to more complex configurations;
• A neural network tool, trained on CFD results, aims to replace CFD simulation capacity by representing impact surfaces with sufficient accuracy. The AI tool's results can be validated and refined through additional CFD simulations;
• The stochastic approach aims to represent realistic spray conditions by modeling the spatio-temporal distribution of particles, including their radius, velocity, temperature, and impact point;
• The combination of the stochastic approach and the AI tool will surpass the capabilities of CFD simulations by enabling the representation of large surfaces and large numbers of particles.

The PhD thesis focuses on the AI/CFD component of the project. Artificial intelligence has been used for several years to detect, segment and reconstruct the contour of the interface between two fluids in experimental images [3, 4]. Rather than segmentation or object detection in an image, the objective of the AI tool to be developed is the prediction of a final state from an initial state and relevant parameters. Given an initial condition consisting of a field of volume fractions representative of the surface state and a characterisation of the particle and its point of impact before impact, the tool must predict the new volume fraction state representative of the droplet's spreading and subsequent solidification. The work will start with a bibliography of available AI analysis tools based on image and volume fraction processing. It will also involve creating a database of droplets impacting substrates with an increasing level of topological complexity, Python programming, comparing the selected methods and 2D/3D verification of the proposed approaches. Particular attention will be paid to quantifying the reliability of predictions [5].

Please send a detailed CV, undergraduate and Master's transcripts, Master's reports and reference contacts to all the following contacts:

Emmanuelle Abisset-Chavanne I2M-Bordeaux : emmanuelle.abisset-chavanne@ensam.eu
Stéphane Glockner I2M-Bordeaux: glockner@bordeaux-inp.fr
Vincent Rat IRCER-Limoges: vincent.rat@unilim.fr
------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Début de la thèse : 01/09/2026

Other public funding

ANR

Université de Bordeaux

209 Sciences Physiques et de l'Ingénieur

Python, Fortran, AI, Computational Fluid Mechanics.
Python, Fortran, AI, Computational Fluid Mechanics.