140 - Integration of neural networks in industrial process control adaptation logic of a run-out table controller for hot strip mills

Who are we

ArcelorMittal is one of the world's leading steel companies. Every day, its 154,000 employees in 60 countries help build a better world with smarter steels that consume less energy and emit less carbon.

The Maizières campus

The Maizières campus is the group's largest R&D site with more than 700 employees who work daily to make steel a cutting-edge product. Coming from the best academic backgrounds, its community of researchers, made up of more than 35 nationalities, designs and industrializes new products for the automotive, packaging and construction markets. To meet the group's ambitions, the decarbonization of steel production is now one of the main drivers of research carried out on the site.

Reference Offre 140

The hot strip mill is the process where a hot-rolled coil is produced from a slab after being rolled multiple times in the rolling stands and then cooled with water using a run-out table to control the shape and the mechanical properties of the strip.
A run-out table is key process in hot rolling composed of water jets controlled by multiple valves to reach a desired temperature along the length of the product. R&D developed an industrial water-cooling controller for run-out table based on temperature and metallurgical predictions.

Temperature control is a key challenge in such manufacturing processes, where product quality strongly depends on the ability to accurately predict and regulate thermal behaviour under a wide range of operating conditions.

Current run out table control systems rely on physical and mathematical models to predict process behaviour and determine optimal operating parameters. While these models provide robust and accurate predictions in most situations, their performance may decrease when operating outside their calibration domain or when confronted with atypical process conditions.

The objective of this internship is to investigate and develop a hybrid modelling approach combining physical models and neural networks to improve the adaptation capabilities of the run out table controller. The work will consist in leveraging historical process data and existing adaptation mechanisms to enhance prediction accuracy and controller performance while maintaining the robustness required for industrial deployment.

The activities of the mission will include:

- Understanding the industrial process and the existing control strategy.
- Reviewing state-of-the-art machine learning approaches for industrial applications.
- Analysing historical process and production data.
- Developing and training neural network models in Python.
- Reusing and extending existing adaptation algorithms.
- Integrating machine learning models into the control adaptation workflow.
- Evaluating model performance through offline simulations and industrial case studies.
- Comparing conventional physical-model-based and hybrid AI-based approaches.
- Documenting developments and presenting results to the R&D team.

The training will be provided by a team of technicians and engineers in Maizières lab. Some contacts with ArcelorMittal Hot Strip Mills could be requested to gather information.

The language of everyday communication is in French but reports and presentations will be in English. The internship offers the opportunity to combine data science, software development and industrial process understanding on a real-world application.

Niveau d'étude: Bac+5 / Master
Domaine: Computer Science / Software Development
Data Science / AI / Image Processing
Mathematics / Statistics
Electronics / Mechatronics / Automation
Modeling / Simulation
Compétences: Python programming skills, Knowedge on Machine Learning and neural network, Analytical and
problem-solving abilities, Ability to communicate technical concepts and present results in a
multidisciplinary environment, Autonomous, curious and willing to learn.
Langues: English, French