PhD Position (f/m/d) in Applied Geosciences: Pattern recognition in DAS data

Area of research

Other

Starting date

Job description

Distributed Acoustic Sensing (DAS) is a fiber optic technology that turns optical fibers into dense seismic arrays. When deployed on unused telecommunication fibers ("dark fibers"), DAS provides regularly spaced seismic measurements along tens or much more kilometers, which could enable seismic monitoring of large areas. The research is planned in the frame of the RUBADO project (BMWE, FKZ 03EE4076A), within which DAS is applied in the Upper Rhine Graben to explore its potential for monitoring geothermal reservoirs and induced seismicity at such scales. Efficient monitoring requires automated processing of the large volumes of data generated (several TB) to extract transient seismic signals, such as microseismic events, from the anthropogenic noise, which constitutes most of the recorded signal. Additionally, identification of quiet periods is of interest for applying ambient seismic noise interferometry. Machine learning (ML) offers a promising solution to automatically classify signals of interest. The objective of this research work is to develop, implement and validate ML-based methods that improve signal detection and classification in DAS data, directly contributing to geothermal monitoring and broader seismic applications.

In this framework, the following tasks are expected:

• Data acquisition, signal pre-processing and classification
• Collect and organize datasets acquired within the RUBADO project
• Perform multi-domain analysis of DAS waveforms in the time, frequency, and space-wavenumber domains (and other array-based representations where relevant)
• Develop robust pre-processing workflows (e.g., denoising and segmentation) tailored to DAS data characteristics
• Identify and extract physically meaningful signal attributes and recurring waveform patterns that capture the variability of seismic and anthropogenic sources, forming the basis for machine learning feature spaces
• Training dataset development and pattern recognition framework
  • Build and curate a labeled dataset through manual inspection and expert annotation of transient signals in DAS recordings
  • Define consistent labeling strategies for different signal classes (e.g., seismic events, traffic-induced noise, instrumental artifacts)
  • Investigate and implement pattern recognition approaches to identify recurrent waveform structures and spatio-temporal signatures in DAS records
  • Develop machine learning and deep learning workflows for automatic signal classification, including supervised, unsupervised, and/or semi-supervised (hybrid) approaches to use both labeled and unlabeled data
• Model validation, benchmarking and transfer
  • Apply ML models to DAS datasets from the RUBADO project
  • Benchmark the performance against independent geophone data and existing event catalogs
  • Assess model generalization capability across different DAS deployments, acquisition geometries, and environmental conditions
  • Perform systematic uncertainty and bias analysis to identify limitations and improve model transferability
• Workflow integration for seismic monitoring and subsurface imaging
  • Integrate the developed processing and machine learning pipeline into the RUBADO analysis framework for near real-time or batch seismic monitoring
  • Enhance event detection, classification, and characterization workflows to improve signal interpretability in DAS data
  • Support improved subsurface imaging by providing cleaner, better-characterized input signals for further seismic processing (e.g., ambient noise analysis, interferometry, or velocity inversion)