Faites connaître cette offre !
Reference : UMR8023-FRELEC-001
Workplace : PARIS 05
Date of publication : Monday, October 07, 2019
Scientific Responsible name : LECHENAULT Frédéric
Type of Contract : PhD Student contract / Thesis offer
Contract Period : 36 months
Start date of the thesis : 1 November 2019
Proportion of work : Full time
Remuneration : 2 135,00 € gross monthly
Description of the thesis topic
We have recently shown that a model knitted fabric exhibits crackling noise in its force response that corresponds to extended fault-like slip bands in its spatial deformation field. The scale invariance displayed by these features, together with their morphology, is strongly reminiscent of seismic events. This system can thus, to some extent, be considered as a toy model for earthquakes.
On the other hand, the emergence of extremely powerful machine learning techniques, mainly represented by elaborate artificial neural networks and so-called “deep learning”, has given some hope to the geophysical community that some aspects of the geo-seismic activity could be inferred from past measurements, the Graal of which being, of course, accurate earthquake prediction. In this case however, the large amounts of data required to train deep nets to be able to predict geophysical activity are very costly to gather and hard to access, and thus far seemingly not sufficient.
On the contrary, in our model setting, we can produce vast amounts of data at virtually zero cost and moderate time. The objective of the internship is thus to collect and analyze “seismic” data from the deformation of a model knit on the one hand, and to try to train a neural networks on the corresponding time series to predict relevant knit-quake quantities. To follow up on this idea, an attention mechanism can be included to the network that can point to what it focuses on to make prediction. Our ultimate goal is to decipher this information to rationalize what in the historic data is relevant to make the prediction, and eventually to compare with geo-seismic data.
The internship is both experimental and theoretical and calls on to various techniques: fabrication and characterization of knitted samples, fine force/displacement measurements, image processing and, on the theoretical side, identifying relevant neural network architectures and input/output formats.
Constraints and risks
We talk about it on Twitter!