Ph.D. proposal - CIFRE funding Physically informed machine learning for manufacturing technologies optimization : from architecture design to online and in-situ implementation
- Paris
- CDD
- Temps-plein
OCTO Technology, a subsidiary of Accenture, is a consulting firm with over 700 employees, including 120 specialists in Data and AI technologies. Serving clients across a wide range of industries, OCTO has established strong partnerships with leading manufacturing companies. Committed to open source, it both utilizes and contributes to open-source software. OCTO fosters an internal culture that encourages growth and continuous learning - employees have access to a comprehensive training catalog covering both foundational and advanced topics, from data science and AI to coding best practices, cloud platforms, green IT, and soft skills. OCTO Technology is proudly certified as a BCorp® and recognized as a Great Place to Work®.Ph.D. project in a nutshell:
The digital transformation of industrial processes relies on sophisticated instrumentation. Machines must be equipped in-situ with smart sensors and supported by systems that can process such as images and time series, in real time. Machine learning and AI have become essential for driving intelligent decisions. In particular, when low latency is critical, these models are deployed directly on devices at the edge, close to the equipment itself - a necessary condition for smarter, faster operations.In general,, decision-taking models can be integrated in two ways:
- In an open-loop fashion, by triggering alerts that are processed by human operators who make the actual decisions;
- Or in a closed loop , by feeding back the decisions into the processes when their criticality allows such unsupervised behavior.Conventional machine learning models suffer from significant limitations in both scenarios:
- They require large amounts of field data for training, and collecting this data can cause delays
- Their memory footprint may not be compatible with edge computing, especially when high precision and low latency are required.
- Their precision often degrades in the field, when the environment is too noisy, or the equipment is subject to regimes not seen during the data collection phase. This issue can have safety implications, particularly in closed-loop setups.Physically Informed Machine Learning (PIML), and in particular Physics-Informed Neural Networks (PINN), are less dependent on data and less energy-intensive, while their predictions remain consistent with physical laws. They offer a promising solution to the triple challenge outlined above. However, their contribution to advanced industrial process control remains under-explored; the current approach relies on statistical tests or conventional machine
learning.One of the manufacturing processes addressed in this thesis is injection molding. In light of environmental concerns, global policies are moving toward greater use of recycled polymers. Reusing polymers that have already been used reduces the carbon footprint of the product. In order to help the automotive industry reduce its carbon emissions, numerous regulations have been enacted to accelerate the use of recycled polymers. However, two major problems remain. First, global production of premium polymers varies greatly depending on the type of polymer. This production volume varies even more widely across industrial sectors. As a result, recycled polymers are made up of elements of different chemical natures. In addition, mechanical recycling, which is currently the most widely used process, causes chain breaks and thus alters the rheological and mechanical behavior of the polymer. For these two reasons, it is essential to monitor the injection process parameters in real time in order to adapt to the rheological behavior of recycled polymers, with the aim of ensuring optimal and repeatable part quality. Initial work has been carried out at PIMM laboratory to add new measuring instruments in order to determine the viscosity of the polymer in real time. The aim now is to use these measurements in addition to the information obtained from the manufacturing process to achieve closed-loop control based on PIML/PINN models.In order to strengthen OCTO's offerings with scientifically grounded innovations, OCTO Technology and the PIMM laboratory at ENSAM are jointly sponsoring this PhD thesis. The research will focus on the application of Physics-Informed Machine Learning (PIML) and Physics-Informed Neural Networks (PINNs) in industrial settings, particularly for edge computing, with the outcomes contributing to open-source software.Ph.D. Objectives:
The objectives are the following;
- Evaluate the eligibility of PIML / PINN models in both open and closed-loop contexts;
- Derive the necessary conditions for that eligibility (data to collect, model architectures, training procedures, …);
- Implement PIML / PINN models on actual equipment, and evaluate their performance against both simulations and field measurements;
- Extend OCTO Technology's open source software ( ) with new process control capabilities, beyond its current specialization in visual inspection.Required skills : If you hold a Master's degree in machine learning, artificial intelligence or applied mathematics with a strong interest in industrial process or you hold a Master's degree related with applied physics or instrumentation and would like to deepen your expertise in machine learning, this Ph.D. offer is for you.Application process: You can apply for this PhD position by sending a CV, a cover letter, your transcripts from the master's degree and/or engineering diploma, and a list of two or three people who can recommend you at , and .Bibliography :
- Cuomo, Salvatore, et al. "Scientific machine learning through physics-informed neural networks: Where we are and what's next." Journal of Scientific Computing 92.3 (2022): 88.
- Karniadakis, George Em, et al. "Physics-informed machine learning." Nature Reviews Physics 3.6 (2021): 422-440.
- Willard, Jared, et al. "Integrating scientific knowledge with machine learning for engineering and environmental systems." ACM Computing Surveys 55.4 (2022): 1-37.
- Karpatne, Anuj, et al. "Theory-guided data science: A new paradigm for scientific discovery from data." IEEE Transactions on knowledge and data engineering 29.10 (2017): 2318-2331.
- Baker, Nathan, et al. Workshop report on basic research needs for scientific machine learning: Core technologies for artificial intelligence. USDOE Office of Science (SC), Washington, DC (United States), 2019.
- A. L. Bowler, M. P. Pound, et N. J. Watson, « A review of ultrasonic sensing and machine learning methods to monitor industrial processes », Ultrasonics, vol. 124, p. 106776, août 2022, doi: 10.1016/j.ultras.2022.106776.
- T. Ageyeva, S. Horváth, et J. G. Kovács, « In-Mold Sensors for Injection Molding: On the Way to Industry 4.0 », Sensors, vol. 19, no 16, p. 3551, août 2019, doi: 10.3390/s19163551.
- Q. Lin, N. Allanic, P. Mousseau, Y. Madec, G. Beau, et R. Deterre, « On-line melt temperature measurements for polymer injection molding through an instrumented annular duct », Polym. Eng. Sci., vol. 62, no 12, p. 3994-4004, 2022.Where to apply E-mailmarc.rebillat@ensam.euRequirementsResearch Field Mathematics » Applied mathematics Education Level Master Degree or equivalentResearch Field Engineering » Process engineering Education Level PhD or equivalentLanguages ENGLISH Level GoodAdditional InformationWork Location(s)Number of offers available 1 Company/Institute PIMM Laboratory Country France GeofieldContact CityParis WebsiteStreet151 Boulevard de l'HopitalSTATUS: EXPIREDShare this page
EURAXESS