Phd offer (CIFRE) - artifical intelligence H/F

C-TEC IS RECRUITING

PhD Offer (CIFRE) 

Constellium is a world leader in the development and manufacture of high value-added aluminum products and solutions for a wide range of markets and applications, focusing in particular on aerospace, automotive and packaging. Our Research and Technology Center, C-TEC Constellium Technology Center employs about 240 people, mainly dedicated to research in the fields of casting, aluminum transformation and surface treatment. We are committed to minimizing the environmental impact of our operations and improving the environmental footprint of aluminum throughout the value chain.

Thesis subject: Artificial intelligence assisted digital twins of hot compression tests on aluminium alloys

Context: Hot forming processes for aluminum alloys play a central role in many industrial sectors, particularly in the manufacture of rolled products. These processes involve severe thermomechanical stresses, combining large deformations, high deformation speeds, and significant temperature gradients. Mastering these operations requires detailed knowledge of the elastoviscoplastic behavior of materials and their damage mechanisms, which are highly dependent on thermal conditions and stress states.

In order to characterize these behaviors under conditions representative of industrial processes, Constellium has developed an innovative hot compression test that analyzes both compression behavior and induced tensile damage mechanisms. This test is used to identify the behavior laws of aluminum alloys, to study their damage/fracture, and to prototype forming processes, particularly rolling. However, the full scientific exploitation of these tests remains limited. The analytical models currently in use are based on simplifying assumptions and empirical corrections, while experimental access to internal deformation, stress, and temperature fields remains limited, particularly for thick specimens and at high temperatures.

Furthermore, the numerical modeling of these tests raises specific difficulties related to large deformations, thermomechanical coupling, and the localization of fields during damage. Although the finite element method allows these phenomena to be described in detail, its computational cost and the complexity of the models limit its direct use for the systematic analysis of tests and for the rapid identification of material laws. In this context, physics-informed artificial intelligence methods are a particularly relevant lever. Based on reference finite element simulations and experimental and industrial data, these approaches make it possible to build interpretable reduced models that can speed up calculations while maintaining the thermomechanical consistency of physical models. AI thus offers the possibility of better exploiting partial or heterogeneous data, assisting in the identification of behavior and damage laws, and making digital twins of hot compression tests truly usable in an industrial setting. It is used here as a complementary tool to finite element modeling, to aid in understanding material mechanisms and optimizing processes.

Objective: The main objective of this thesis is to develop a thermomechanical digital twin of hot compression tests on aluminium alloys, based on finite element simulations and enriched by machine learning methods integrating physical knowledge. This digital twin aims to improve the scientific and industrial exploitation of hot compression tests, adapted to the thermomechanical loading paths of hot rolling. It will initially enable rapid and robust identification of viscoplastic rheological laws, then damage laws, under thermomechanical conditions representative of industrial forming. Unlike conventional approaches, based on simplified analytical models or heavy experimental campaigns, the thesis aims to fully exploit the wealth of three-dimensional stress, strain, and temperature fields provided by finite element simulations, while combining them with experimental observables—forces, displacements, temperature measurements, or fracture surfaces—in order to make accelerated models both more reliable and more representative of actual tests.

Methods: The innovative nature of the project is largely based on the methodology adopted. The thesis is part of the field of scientific machine learning, which explicitly integrates knowledge from materials mechanics into learning models. Far from “black box” approaches, the models developed are “gray box” models, combining the rigor of reference finite element simulations (LS-DYNA) with the efficiency of reduced models trained on multimodal data from tests, simulations, and existing material databases. An original two-step strategy is implemented: pre-trained models learn, from large simulated databases, latent representations of thermomechanical fields under large deformations; these models are then adapted to specific application tasks (calibration, rapid prediction, test design) based on a limited amount of experimental data.

This approach overcomes one of the major obstacles to learning in mechanics: the scarcity of rich, labelled experimental data.

The thesis also explores advanced representations of mechanical fields, particularly in the form of point clouds, adapted to large deformations and the localization of fields during damage. The use of modern model reduction techniques and autoencoders opens up new perspectives for the accelerated modelling of complex thermomechanical phenomena, which have yet to be explored in the context of metal forming.

We are looking for a highly-motivated individual with a Master’s degree or Engineering degree in materials mechanics or mechanical engineering with strong skills in machine learning, particularly for models incorporating physical knowledge (scientific machine learning, autoencoders, reduced models). Experience in finite element modelling and complex multi-physics data exploitation will be highly appreciated. The candidate must demonstrate autonomy, scientific rigor, and curiosity for the development of thermomechanical digital twins and digital prototyping of processes.

Starting date: October/November 2026              Duration: 3 years

Host laboratories: CEMEF (80-90%) and Constellium C-TEC (10-20% with possible adjustment during the different stages of the thesis)

Academic supervision (CEMEF): David Ryckelynck, Pierre-Olivier Bouchard

Industrial supervision (Constellium C-TEC): Alexandre Barthelemy