CE10 - Industrie et usine du futur : Homme, organisation, technologies

Knowledge and artificial intelligence aided Additive Manufacturing – KAM4AM

KAM4AM

In the field of additive manufacturing, generating multi-axis material deposition trajectories remains a particularly strong challenge for DED-type processes. KAM4AM aims to develop an innovative software solution for manufacturing based on a proven artificial intelligence (AI) technique, i.e. Reinforcement Learning (RL), to make it a learning and adaptive CAM.

A learning CAM for DED

While not dealing with fundamental research on Artificial Intelligence algorithms, the KAM4AM project proposes to develop, within the framework of a particular field of application related to additive manufacturing, a concrete contribution to Knowledge Engineering, under the form of a process of formalization of expert knowledge to lead to their mathematization and their introduction into a learning agent-based system for the generation of DED trajectories.

The project follows an approach based on knowledge engineering, on the one hand, and on artificial intelligence on the other. Hence, elicitation of knowledge from DED manufacturing experts allows the establishment of a powerful reward system for the adopted reinforcement learning algorithms. Coupled with finite element thermomechanical simulations, this expert knowledge is also mobilized for the development of rapid phenomenological models to simulate the deposition of material during manufacturing. This simulation is necessary for the evaluation of manufacturing trajectories during the multiple reinforcement learning iterations.

The final objective of the project is the creation of a CAM system based on knowledge and artificial intelligence, dedicated to additive manufacturing DED. The first results of the project concern several bricks of this system:
- a first version of a learning environment, based on a modular software architecture, has been created. The Gym library was chosen for reinforcement learning algorithms. This allows the use of all standard algorithms and guarantees compatibility with future algorithms. The proposed environment has been tested in the case of simple learning. For this, a simplistic simulator based on the voxelization of the manufacturing space has been created. This makes it possible to simulate the manufacture of a part and to check its correct geometric filling. The learning agent can thus test different actions such as moving or switching on the energy source. The studies carried out have shown a good convergence of the learning agent and a possibility of generalization to other types of parts than the parts used for learning. In addition, these results made it possible to characterize the number of actions necessary for good learning and therefore the tolerable computing time (of the simulator) for learning in a reasonable time. The following work was devoted to obtaining a fast simulator. Finally, the key role of the reward system was highlighted. This will be the subject of future studies.
- A version of the finite element thermal model of the WAAM manufacturing process (high-fidelity 3D model concerning the generation of a wall) was created and validated against the results available in the literature. A database for the variation of different deposition parameters (Travel Speed, Torch Power and waiting time between the deposition of two layers) was built. This base is necessary to feed the NURBS-based phenomenological meta-model that will be created later.
- a real-time thermal simulator of the temperatures on the upper layer of the substrate has also been developed. Less ambitious in terms of precision than the final NURBS-based phenomenological simulator should be, it will allow the further development of the learning environment and the associated reward system.

The project will continue to work on the phenomenological thermal simulator, which will in particular require carrying out an experimental study based on the manufacturing of several test parts and numerous calculations using the proposed finite element model. There are also plans to develop a rapid geometric simulator to assess the shape of the bead deposited based on the simulated temperature map. At the same time, the work will also concern the formalization of expert knowledge and the construction of the reward system that will be based on it.

Tezenas Du Montcel, T., Beraud, N., Vignat, F., Pailhès, J., Marin, P., Pourroy, F. (2023). Real-Time Approximative Thermal Simulation for Wire Arc Additive Manufacturing. In: Gerbino, S., Lanzotti, A., Martorelli, M., Mirálbes Buil, R., Rizzi, C., Roucoules, L. (eds) Advances on Mechanics, Design Engineering and Manufacturing IV. JCM 2022. Lecture Notes in Mechanical Engineering. Springer, Cham. doi.org/10.1007/978-3-031-15928-2_50
Vignat, F., Béraud, N., Montcel, T.T.D. (2023). Toolpath Calculation Using Reinforcement Learning in Machining. In: Gerbino, S., Lanzotti, A., Martorelli, M., Mirálbes Buil, R., Rizzi, C., Roucoules, L. (eds) Advances on Mechanics, Design Engineering and Manufacturing IV. JCM 2022. Lecture Notes in Mechanical Engineering. Springer, Cham. doi.org/10.1007/978-3-031-15928-2_100
M. Zani, B. Vuillod, M. Montemurro, E. Panettieri, P. Marin. A metamodel based on NURBS hypersurfaces to simulate the wire arc additive manufacturing. 2nd International Conference on Computations for Science and Engineering ICCSE2 Rimini Riviera, 30 August - 2 September, 2022

In Additive Manufacturing, Directed Energy Deposition (DED) is a promising technology that gains a growing interest in industry. An essential feature of this process its rapid fabrication capability, even for large-size parts. However, generating good material deposition trajectories remain a huge challenge that CAM software often fail to correctly deal with. The KAM4AM project aims at developing a software for DED manufacturing, based on the proven Artificial Intelligence technology of Reinforced Learning, to get a learning and adaptive CAM solution. A list of study cases from industry will help to collect the typologies of parts as well as technical and scientific issues related to DED technology. This data, combined with research cases, will enable to define the objectives and the functions of the learning environment that needs to be created.
The main research challenges are (1) to design a problem-independent reward system, based on expert rules of the DED domain, (2) to develop a phenomenological model of the DED process, fast enough for allowing the numerous iterations required for the learning process. A last step consists in a thorough test of the generated trajectories, followed by the integration of these trajectories into Esprit Additive software.

Project coordination

Franck POURROY (Laboratoire des Sciences pour la Conception, l'Optimisation et la Production de Grenoble)

The author of this summary is the project coordinator, who is responsible for the content of this summary. The ANR declines any responsibility as for its contents.

Partner

G-SCOP Laboratoire des Sciences pour la Conception, l'Optimisation et la Production de Grenoble
DPRI DP Research Institute / R&D
I2M INSTITUT DE MECANIQUE ET D'INGENIERIE DE BORDEAUX

Help of the ANR 549,652 euros
Beginning and duration of the scientific project: - 48 Months

Useful links

Explorez notre base de projets financés

 

 

ANR makes available its datasets on funded projects, click here to find more.

Sign up for the latest news:
Subscribe to our newsletter