Analysis of continuous cryo-seismic time series with deep learning for future in-situ analysis & smart sensor on sea ice.

Durations : 5-6 months ; dates to be determined.

Scientific context

The Arctic is changing rapidly, and understanding how sea ice responds to environmental forcing is becoming increasingly important. In the Marginal Ice Zone, ocean waves interact strongly with sea ice, generating complex seismic signals that contain information about ice dynamics and mechanical properties. These signals include wave propagation, icequakes and other sources of seismic energy, whose characteristics depend on factors such as ice thickness and mechanical properties.

Seismic methods offer a unique opportunity to monitor these processes at fine spatial and temporal scales. However, extracting relevant information from seismic recordings remains challenging, particularly when working with continuous data, where signals of interest are embedded in complex and variable background noise. Furthermore, training deep-learning models generally requires large amounts of labelled data, which are often difficult and costly to obtain in remote environments such as the Arctic.

Recent advances in self-supervised learning offer promising ways to exploit large amounts of unlabelled data and improve the performance of neural networks on downstream tasks.

Objectives and tasks during the internship

The objective of this internship is to investigate how continuous seismic recordings acquired by geophones on sea ice can be exploited to learn representations of the underlying physical processes. The focus will be on convolutional neural networks (CNNs) and self-supervised or unsupervised learning strategies, with the aim of improving the estimation of sea-ice properties while limiting the need for labelled training data.

The internship will involve the following tasks :

• Pre-training the network on the continuous time series with strategies such as self-supervised learning and contrastive learning,
• Measure the impact of such strategies on the existing architectures for ice-quake distance estimation and ice thickness estimation,
• Evaluate the approaches in terms of computational ressources. In particular the end goal of this project is to deploy the algorithm within the acquisition sensor

The internship will build on existing datasets and neural-network architectures developed within our research group. It is closely connected to an ongoing PhD project on deep learning for sea-ice mechanical-property estimation and complements another internship focused on the detection and location of icequakes using a two-step approach. In contrast, this project will investigate whether useful information can be extracted directly from continuous recordings, potentially paving the way for autonomous, real-time analysis in the field.

The project lies at the intersection of geophysics, seismology, Arctic science and artificial intelligence. It offers the opportunity to work with real-world data and explore modern machine-learning methods for environmental monitoring.

This internship is part of the ANR MSIM project (Multiscale Sea Ice Imaging) and is connected to the MIAI chair INSITU (Embedded Artificial Intelligence for Environmental Monitoring).

Profile and skills

• Python programming
• Signal processing : Fourier analysis, time-frequency analysis, cross-correlation, etc.
• Basic understanding of machine learning and deep learning, with some practical experience
• Background in geophysics or seismology is an advantage

To apply, please send an email with your CV and a short statement of motivation.