Détermination des propriétés mécaniques de profilés extrudés en aluminium par intelligence artificielle

Abstract

The main goal of this project is to propose a simple model to calculate the mechanical properties and the anisotropy of extruded aluminum profiles using artificial intelligence (AI). AI algorithms work as a statistical approach which could help in reaching better predictions than the ones given by heavy calculations deployed in traditional texture evolution models used to simulate extrusion. Particular microstructural transformations are known to occur within the material when passing through the extrusion die, which heavily modifies the crystallographic texture and creates strong anisotropy in the extruded profiles. Their mechanical and microstructural properties were studied in the current project for two commercially available alloys, known as AA6063 and AA6082, often used in the transportation industry. The microstructure morphology, the cristallographic texture and the precipitation state were studied for various extrusion parameters and various beam geometries, in order to relate them with the mechanical properties. Alloy AA6063 exhibited a recrystallized microstructure primarily with big equiaxed grains. Its texture is dominated by randomlyoriented grains and cube component, and seemed mainly driven by deformation path and billet pre-heating temperature. This alloy mechanical properties revealed to be mainly sensitive to ram speed, and the Lankford coefficient, being representative of anisotropy, has primarily been influenced by its crystallographic texture. Alloy AA6082 exhibited a deformed microstructure with small fine grains. Its texture is dominated with S and Brass texture components, which seemed particularly sensitive to the extrusion speed and path. Its mechanical properties didn’t seem being really influenced by extrusion parameters on their studied range. Crystallographic texture measurements were used to apply the CMTP method, giving the possibility to predict the Lankford coefficient with rather good tendancies, but a lack of precision for AA6063 profiles which contain much more randomly oriented grains than AA6082 profiles. Extrusion data and mechanical properties were used to train multilayer perceptrons (MLP) and support vector regressions (SVR), each of these predicting one of the following quantities: yield and ultimate strengths, hardening coefficient and Lankford coefficient. The results calculated by MLPs always were better than SVRs, and they were only satisfying for ultimate strength, but disappointing for other quantities.

Date	24 Aug 2026
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Philippe Bocher (Supervisor)