Propositions de sujets de these
Learning anomaly detection in nanoindentation to identify crystal elastoplasticity properties at the sub-grain level
Learning anomaly detection in nanoindentation to identify crystal elastoplasticity properties at the sub-grain level
Mots-clés
Machine learning, Anomaly detection, plasticité, Équations constitutives
Machine learning, Anomaly detection, Plasticity, Constitutive equations
Résumé
/
Description of the PhD project :
The present PhD project will focus on anomaly detection and images-based machine learning to correlate hardness and reduced modulus of oxidised and non-oxidised Ni-based and Ti-based materials as a function of the grain orientation, the local chemical composition and metallurgical state as well as the proximity of grain boundaries. High resolution nanoindentation maps either in a conventional mode or a continuous stiffness measurement mode will be performed to assess the local mechanical properties.
In the present project, the PhD student will:
- Perform nanoindentation tests on model materials to get more experimental data;
- Data augmentation by using a pre-existing crystal plasticity finite element model with strain gradient plasticity simulating the nanoindentation test in a single crystal;
- Develop a machine learning code sampling multimodal data to fastly predict the nanoindentation properties as a function of grain orientation, chemical composition and the metallurgical state;
- To detect surface events/anomalies from surface observation related to the nanoindentation response;
- Cluster data to identify set of microstructural features favouring particular mechanical response (slip activity, hardness-elastic response relationship, etc.);
- Identify parameters of the crystal plasticity finite element model using machine learning;
Contexte
What is a normal data in science of materials? Structural metallic/intermetallic materials
operating at high temperatures (650°C-1200°C) in severe environments are commonly subjected to inservice surface reactivity, i.e., oxidation, corrosion. This issue is encountered in several industrial applications, especially when high temperatures, mechanical stresses, and highly corrosive atmospheres gather (power plants, aeronautic turbines, etc.) [1]. In this framework, surfaces of mechanical components are not uniform. The gradient of microstructure and the related variability in mechanical behaviour often drives premature damage and the progressive rupture of the components [2].
Automated surface-anomaly detection using machine learning has become an interesting and promising area of research, with a very high and direct impact on the application domain of visual inspection. Deep-learning methods have become the most suitable approaches for this task [3]. Convolutional neural methods have proven to outperform other approaches in various computer vision tasks, such as automated visual surface inspection [4,5]. In industrial applications with well-optimized processes is it often not possible to obtain a sufficiently large set of defective samples for CNN-based classification [6]. Hence, anomaly detection is usually trained on normal data [7,8]. A distillationanomaly detection model has been proposed in [9] to improve the detection of abnormal samples. Various architectures of neural networks have been proposed in the literature, such as autoencoder [8] and transformers [Lu2022].
The data set for this work is obtained by in-lab fast scanning nanoindentation and high resolution-digital image correlation (HR-DIC). These are advanced characterisation techniques giving access to local mechanical properties at the sub-grain level. These techniques, correlated to additional chemical or orientation analyses, generate a considerable volume of data compatible with artificial intelligence techniques to cluster and analyse data from different formats (images, vectors, scalars, curves, etc.). These observational data are labelled as normal data or abnormal data, prior further analysis of normal data only. But this labelling step is time consuming because it requires an expert knowledge.
The present project intends to better assess the local mechanical behaviour at the sub-grain level using more normal data from automated surface inspection and nano-indentation techniques in Ti- and Ni-based materials. Normal observational data will be augmented by numerical simulations of the nanoindentation test in single crystals. This numerical step aims to perform an hybrid artificial intelligence that does not ignore the scientific knowledge that have been incorporated in physic-based numerical models. Such numerical simulations aim at performing a physics-guided machine learning of a category of normal data. Simulated data and observational data are multimodal data in huge dimensional spaces that contain hidden empirical information that should be summarised in a smaller latent space of normal data. Such a latent space can be learnt by an autoencoder [7]. Data significantly distant from this latent space will be labeled as abnormal data.
Project motivations :
To address this point, BigMeca chair and HT-S4DefOx project join together in order:
- To automate acquisition of normal data in fast scanning nanoindentation;
- To develop a comprehensive methodology combining experimentation, simulation and machine learning to emulate physics-based predictions for normal data only;
- To detect abnormal observational data in fast scanning nanoindentation;
- To assess the mechanical behaviour within the time-evolving gradient of microstructure and properties, i.e. within the “sub-surface” material (micro- and mesoscale approach);
- To assess the variability in sub-grain mechanical behaviour of the metallic material at the metal/oxide interface (microscale approach). This interface, considered as the “extreme surface”, is on the front line for the thermo-mechano-chemical coupling;
- To apply statistical learning in solid mechanics to better understand elementary deformation mechanisms, and crystal elastoplasticity properties.
Encadrement
Directeur de thèse : David RYCKELYNCK - Centre des Matériaux
Co-Directeur de thèse : Damien TEXIER - CNRS
Co-encadrant : Henry PROUDHON - Centre des Matériaux
Profil du candidat
Profil type pour une thèse à MINES ParisTech: Ingénieur et/ou Master recherche - Bon niveau de culture générale et scientifique. Bon niveau de pratique du français et de l'anglais (niveau B2 ou équivalent minimum). Bonnes capacités d'analyse, de synthèse, d'innovation et de communication. Qualités d'adaptabilité et de créativité. Capacités pédagogiques. Motivation pour l'activité de recherche. Projet professionnel cohérent.
Pré-requis (compétences spécifiques pour cette thèse) :
The PhD student should have the following skills and/or know-how:
- Scientific computing (Python language, etc.);
- Mechanical engineering (and more particularly at the microscale);
- Artificial Intelligence, machine learning.
Pour postuler : Envoyer votre dossier à recrutement_these@mat.mines-paristech.fr comportant
- un curriculum vitae détaillé
- une copie de la carte d'identité ou passeport
- une lettre de motivation/projet personnel
- des relevés de notes L3, M1, M2
- 2 lettres de recommandation
- les noms et les coordonnées d'au moins deux personnes pouvant être contactées pour recommandation
- une attestation de niveau d'anglais
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):
Expected skills and/or know-how :
The PhD student should have the following skills and/or know-how:
- Scientific computing (Python language, etc.);
- Mechanical engineering (and more particularly at the microscale);
- Material sciences and/or computational solid mechanics.
Applicants should supply the following :
- a detailed resume
- a copy of the identity card or passport
- a covering letter explaining the applicant's 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 recrutement_these@mat.mines-paristech.fr
Références
References :
1. Young DJ (2016) High temperature oxidation and corrosion of metals, 2nd Ed. Elsevier Science
2. Pineau A, Antolovich SD (2009) High temperature fatigue of nickel-base superalloys - A review with special emphasis on deformation modes and oxidation. Eng Fail Anal 16:2668–2697. doi:10.1016/j.engfailanal.2009.01.010
3. Tabernik, Domen and Šela, Samo and Skvarč, Jure and Skočaj, Danijel}, Segmentation-based deep-learning approach for surfacedefect detection, 2020, doi:10.1007/s10845-019-01476-x
4. Racki, Domen and Tomazevic, Dejan and Skocaj, Danijel, A Compact Convolutional Neural Network for Textured Surface Anomaly Detection, (2018), doi:10.1109/WACV.2018.00150
5. Vengaloor, Reshma P. and Muralidhar, Roopa, Deep Learning Based Semantic Segmentation Technique for Anomaly Detection on Metal Surfaces Using High Calibre U- Shaped Network, (2022), doi:10.18280/ts.390614
6. Staar, Benjamin and Lütjen, Michael and Freitag, Michael, Anomaly detection with convolutional neural networks for industrial surface inspection, (2029), doi:10.1016/j.procir.2019.02.123
7. C. Huang, F. Ye, J. Cao, M. Li, Y. Zhang, et C. Lu, « Attribute Restoration Framework for Anomaly Detection », arXiv:1911.10676 [cs], déc. 2020, http://arxiv.org/abs/1911.10676
8. Schmedemann, Ole and Miotke, Maximilian and Kähler, Falko and Schüppstuhl, Thorsten, Deep Anomaly Detection for Endoscopic Inspection of Cast Iron Parts, (2023), doi:10.1007/978-3-031-18326-3_9
9. Zhou, Qunying and Wang, Hongyuan and Tang, Ying and Wang, Yang, Defect Detection Method Based on Knowledge Distillation, (2023), doi:10.1109/ACCESS.2023.3252910
10. Lu, Xiaofeng and Fan, Wentao, Transformer-based Encoder-Decoder Model for Surface Defect Detection, (2022), doi:10.1145/3529466.3529471