PhD opportunities

SUBJECT OF THESIS PROVIDED - Machine learning for defect acceptability by using reduced-order modeling, Z-ROM Z-learn

Thesis proposal

Area of expertiseMechanics
Doctoral SchoolSMI - Sciences des Mtiers de l'Ingnieur
SupervisorM. David RYCKELYNCK
Co-supervisorM. Jacques BESSON
Research unitCentre of materials
KeywordsSimulation-driven machine learning, dictionary of defects, multi-scale modeling, aeronautics, energy, HPC
AbstractWe propose to develop a machine learning approach to model order reduction method for the analysis of the harmfulness of defects in materials mechanics. Machine learning algorithms have recently been proposed for defect analysis, for pipelines [1], in aeronautics [2] or in manufacturing processes [3]. In addition, numerical simulation and machine learning have shown their complementarity, for example for the study of defects in bearings [4]. The 'Simulation-driven machine learning' approach [4] is very attractive when models are available. In addition, the machine learning methods make it possible to avoid the parameter setting of the objects to be modeled. This is a particularly interesting property for defect analysis. They will be represented using images (2D or 3D).
In this thesis we will focus on local defects in ductile metallic materials. The objective of the thesis is to constitute dictionaries of numerical models allowing a quick decision on the harmfulness of a local defect. This type of decision takes place during the production phase or during the operation phase of mechanical components. The dictionaries will cover wide variations in behavior or loading parameters and also a wide variety of different geometries.
It is proposed to constitute a first dictionary of mechanical models of sound parts, without defects, on a macroscopic or mesoscopic scale. The models in this dictionary will be hyper-reduced models from a method proposed in [5]. This method has recently been extended, in order to be able to insert local defects in a defect-free model. It allows the study of harmfulness by considerably reducing the number of mechanical equations to solve (several orders of magnitude). In this approach, the defect insertion passes through a reduced model of the defect, considered as isolated in an infinite environment. The second dictionary that is proposed to be developed is that of reduced models of isolated defects in an infinite environment. Experimental 2D or 3D data (from X-ray tomography) will be considered to treat realistic defects. The two dictionaries thus constituted will make it possible to give more values to the experimental data and to the simulation data, which are generated during the studies of harmfulness of defect.
The analysis of defects in infinite medium will be treated by combining the finite element method and the FFT method. The first captures stress gradients very well and the second is very fast to execute. Particular attention will be paid to the generality of the algorithms developed, so that they can be quickly reused by researchers in materials mechanics. In addition, a procedure will be required to enrich the single defect database and the defect-free model database.

[1] Mohamed Layouni, Mohamed Salah Hamdi and Sofiane Tahar, Detection and sizing of metal-loss defects in oil and gas pipelines using pattern-adapted wavelets and machine learning, Applied Soft Computing, (2017),
[2] Biagio, Marco San Beltrn-Gonzlez, Carlos, Giunta Salvatore, Bue Alessio Del and Murino Vittorio, Automatic inspection of aeronautic components, Machine Vision and Applications, (2017),
[3] Carlos A Escobar and Ruben Morales-Menendez, Machine learning techniques for quality control in high conformance manufacturing environment, Advances in Mechanical Engineering, (2018),
[4] Cameron Sobie, Carina Freitas and Mike Nicolai, Simulation-driven machine learning: Bearing fault classification, Mechanical Systems and Signal Processing, (2018),
[5] D. Ryckelynck, K. Lampoh, S. Quilici, Hyper-reduced predictions for lifetime assessment of elasto-plastic structures, Meccanica, (2015), DOI 10.1007/s11012-015-0244-7

Thesis program:
- Preliminary study of a case of a cracked pipe by using the hyper-reduction method available in the research version of the Z-set software developed at the Materials Center.
- Implementation of a digital environment for the development of dictionaries for Z-set.
- Coupling of the FFT method and the finite element method for the insertion of defects in hyper-reduced Z-set models
- Realization of a very large number of simulations by HPC for learning dictionaries.
- Development of a method for using dictionaries by automatic learning. The software developed will be called Z-learn. It will be open-source. It will be written in python to benefit from the many libraries available for automatic learning (Tensor Flow, Keras, ...).
- Development of a partial and mesoscopic regularization method for the study of localized damage.
ProfileTypical profile for a thesis at MINES ParisTech: Engineer and / or Master of Science - Good level of general and scientific culture. Good level of knowledge of French (B2 level in french is required) and English. (B2 level in english is required) Good analytical, synthesis, innovation and communication skills. Qualities of adaptability and creativity. Teaching skills. Motivation for research activity. Coherent professional project.
Prerequisite (specific skills for this thesis): applied mathematics, machine learning or computational mechanics

Applicants should supply the following :
a detailed resume
a covering letter explaining the applicants motivation for the position
detailed exam results
two references : the name and contact details of at least two people who could be contacted
to provide an appreciation of the candidate
Your notes of M1, M2
level of English equivalent TOEIC

to be sent to
FundingFinancement d'un Etablissement d'enseignement suprieur