Microstructural characterization by bulk laser-ultrasounds tomography
The proposed thesis falls into the framework of designing innovative methods in the materials characterization. The thesis aims to develop a new tomographic technique for characterizing microstructures using bulk laser ultrasound. In the state of the art, acoustic methods such as the scanning acoustic microscope and surface wave optoacoustic spectroscopy lead to grain imaging but only at the surface of the sample. However, industrial manufacturing processes (in metallurgy, welding, additive manufacturing...) can reveal spatial inhomogeneity of the microstructure, such as grain size gradients with depth within the component. Electron backscatter diffraction (EBSD) also provides surface imaging of grains but has disadvantages, including restriction on sample size and the need to make transversal cuts through the sample to image its volume.
The proposed idea is to develop a bulk laser ultrasound tomography technique able to determine the grain size in areas of a component or even to obtain imaging of large-grain microstructures and a local determination of crystallographic orientation. Therefore, the thesis's main objective will be to design such an experimental characterization tool and optimize its design using a digital twin to develop.
Conditional generative model for dose calculation in radiotherapy
Particle propagation through matter by Monte Carlo method (MC) is known for its accuracy but is sometimes limited in its applications due to its cost in computing resources and time. This limitation is all the more important for dose calculation in radiotherapy since a specific configuration for each patient must be simulated, which hinders its use in clinical routine.
The objective of this thesis is to allow an accelerated and thrifty dose calculation by training a conditional generative model to replace a set of phase space files (PSF), whose architecture will be determined according to the specificities of the problem (GAN, VAE, diffusion models, normalizing flows, etc.). In addition to the acceleration, the technique would produce an important gain in efficiency by reducing the number of particles to be simulated, both in the learning phase and in the generation of particles for the dose calculation (model's frugality).
We propose the following method:
- First, for the fixed parts of the linear accelerator, the use of a conditional generative model would replace the storage of the simulated particles in a PSF, whose data volume is particularly large. The compactness of the model would limit the exchanges between the computing units without the need for a specific storage infrastructure.
- In a second step, this approach will be extended to the final collimation whose complexity, due to the multiplicity of possible geometrical configurations, can be overcome using the model of the first step. A second conditional generative model will be trained to estimate the particle distribution for any configuration from a reduced number of simulated particles.
The last part of the thesis will consist in taking advantage of the gain in computational efficiency to tackle the inverse problem, i.e. optimising the treatment plan for a given patient from a contoured CT image of the patient and a dose prescription.
Natural language interactions for anomaly detection in mono and multi-variate time series using fondation models and retrieval augmented generation
Anomaly detection in mono and multi-variate time series highly depends on the context of the task. State-of-the-art approaches rely usually on two main approaches: first extensive data acquisition is sought to train artificial intelligence models such as auto-encoders, able to learn useful latent reprensations able to isolate abnormality from expected system behaviors; a second approach consists in careful features construction based on a combination of expert knowledge and artificial intelligence expert to isolate anomalies from normal behaviors using limited examples. An extensive analysis of the literature shows that anomaly detection refer to an ambiguous definition, because a given pattern in time series could appear as normal or abnormal depending on the application domain and the immediate context within the successive observed data points. Fondation models and retrieval-augmented generation has the potential to substantially modify anomaly detection approaches. The rationale is that domain expert, through natural language interactions, could be able to specify system behavior normality and/or abnormality, and a joint indexing of state-of-the-art literature and time series embedding could guide this domain expert to define a carefully crafted algorithm.
High speed/High capacity distributed Fiber Bragg Grating sensing technique for Structural Health Monitoring (SHM) applications
Schedule-driven Non-Destructive Evaluations (NDE) are carried out during structure/equipment’ life to detect major degradations endangering safety and impairing service availability. In addition to NDE, Structural Health Monitoring (SHM) involves the use of in-situ Fiber Bragg Grating (FBG) sensing systems and algorithms to evaluate structure worthiness. FBGs are mostly used as strain/temperature sensors but are also used for acoustic sensing, as substitutes to piezoelectric actuators. The SHM of large structures or acoustic measurements for passive/active tomographic techniques simultaneously require a high capacity and readout rate. However, commercially available FBG readout units rely upon Wavelength-Division Multiplexing (WDM) or Optically Frequency-Domain Reflectometry (OFDR) techniques. WDM-based units are limited in capacity (several tens of sensors) but may reach high scan rate (MHz or beyond MHz). Conversely, OFDR-based units are limited in scan rate (typically several tens of Hz) but may accommodate large number of sensors (typically up to 2000). Tomography with acoustic techniques requires both high capacity and high scan rate with the aim to improve quality of image reconstruction. Optical Time-Stretch (OTS) is a time-domain technique that has potential to improve both capacity and scan rate and to open the way to efficient tomography reconstruction processes. The basics of OTS is to use a pulsed laser, a highly dispersive medium and a high bandpass photodetector in order to convert a Bragg wavelength shift into a time delay. The doctoral candidate will investigate several ways to implement OTS to SHM. Draw-Tower Gratings (DTG) and chirped gratings will be used for the measurement of strain profiles and acoustic field emission on metallic and carbon fiber-reinforced plastics (CFRP) composite structures. The candidate will first assess the performance of the OTS technique in laboratory (LSPM) with piezoelectric actuators and laser-ultrasonics (if available, with CNRS/PIMM). Then, the OTS device will be tested onto several demonstrators provided by partners within the MSCA USES 2 doctoral network: civil engineering structure (BAM, Berlin), hydrogen storage canister (Faber, Cividale del Friuli) and CEA DAM (Le Ripault) and finally onto a metallic pipeline for fluid transport (ENI, Milano). The doctoral candidate will move onto those test sites during three 2-month periods. He will implement the OTS technique and gather experimental feedback.
Study of inversion methods based on simulation and machine learning for defect characterisation in ultrasonic array imaging
The thesis work is part of the activities of CEA-List department dedicated to Non-Destructive Testing (NDT), and aims to study simulation-based inversion methods to characterise defects from ultrasonic images, such as TFM (Total Focusing Method) or PWI (Plane Wave Imaging) images. The inversion methodology will rely on machine learning algorithms and numerical training databases generated with the CIVA software platform. A first part will study the ability of such an inversion method to characterise a defect (location, size, orientation...) without any a priori information, by exploiting the noise and reconstruction artefacts due to the use of unsuitable propagation modes. In a second part, the simulation-based inversion will be evaluated in more realistic situations where images are of poor quality due to uncertainties on the properties of the component and/or on the experimental setup. In order to reduce the generation time of the training database, and to gain in robustness and accuracy, the feasibility of inverting fast imaging (e.g.: combining PWI and fast reconstruction algorithms in the Fourier domain) will be studied, as well as the feasibility of directly inverting signals or spectra without the need to compute images. The inversion method will be experimentally evaluated with different mock-ups representative of industrial components and, at the end of the thesis, a real-time proof of concept will be demonstrated by implementing the imaging and inversion algorithms in a laboratory prototype system.