Deep Learning Tools for Seismic Applications

Deep learning is a groundbreaking technology also for the earth sciences. The project Deep Learning for Large Seismic Applications (DLseis) deals with basic research up to ready-to-use tools.

Project DLseis – Deep Learning for Large Scale Seismic Applications

The availability of large data sets and advances in computing power established Deep Learning (DL) as breakthrough technology for many of the hardest problems in speech, vision and natural language processing. The successful, learned approaches outperform their human-engineered counterparts in accuracy, throughput and flexibility. Many grand challenges in the geosciences share similar characteristics: A strong desire to gain scientific insight from a rapidly growing amount of available synthetic, observational and experimental data.

The interdisciplinary project Deep Learning for large scale seismic applications (DLseis) systematically tackles these challenges for a collection of demanding problems in seismic processing.

The scope of DLseis is three-fold, covering

  • basic research in deep learning methods
  • education and dissemination
  • ready-to-use software for production.

Application-Driven Research

Our basic research addresses distinct scientific requirements before Deep Learning can be applied readily and routinely to seismic applications. First, seismic pre-stack data sets are often high-dimensional and typically exceed even the size of the largest commonly used Deep Learning benchmarks. Thus, DLseis develops the required technology for scalable and high-performance Deep Learning. In addition, seismic field data is often unlabeled which prohibits the direct application of established DL workflows. DLseis mitigates this lack of labeled field data through transfer learning from systematically generated synthetics and by studying unsupervised approaches.

Production-Ready Deep Learning Tools

A collection of demanding key applications in seismic processing transfers our basic research to production-ready tools.
Together with experts in industry we apply and evaluate Deep Learning for challenges in pre-stack processing and interpretation.

Illustration of our DL-driven method to trim statics.
© Fraunhofer ITWM
Illustration of our DL-driven method to trim statics.
Comparison of peak values in two example gathers taken from field data. We observe an excellent peak preservation after alignment through the DL-driven approach.
© Fraunhofer ITWM
Comparison of peak values in two example gathers taken from field data. We observe an excellent peak preservation after alignment through the DL-driven approach.

An example are trim statics which improve seismic inversion and interpretation through post-migration conditioning.
Here, we study a holistic approach combining domain-knowledge with the power of data-driven DL.
For example, the application of a trained neural network for trim statics in the figure above greatly improves the alignment of the subsurface reflections in the shown seismic gathers.

 

 

Our workflow is split into three major parts: 1) The generation of physics-based synthetic data, 2) pre- and postprocessing, and 3) neural network training. Shown is the flow of the synthetic data during training. The parameters of data-dependent preprocessing steps are solely derived using the unaligned input and then applied to the aligned ground truth. During training we repeatedly use the neural network forward to obtain the aligned result. This allows us to derive errors and optimize the network's weights through backpropagation.

A second application is the automuting of seismic gathers, a pre-processing step which eliminates low quality signals before stacking. Traditionally this step is done manually by domain experts. We are developing a Deep Learning solution for this pixel-wise segmentation which improves not only quality and time-to-solution but also reduces human bias.

Interactive visualization enabling quality control and manual corrections of masks inferred in real-time by a neural network
© Fraunhofer ITWM
Interactive visualization enabling quality control and manual corrections of masks inferred in real-time by a neural network

ALOMA – Software for Seismic Experts

Our Software ALOMA is selected as platform for integrating and executing all tools developed. ALOMA is a parallel runtime environment for seismic applications. It is based on ITWM’s HPC tools, like the auto-parallelizing runtime engine GPI-Space and the distributed file system BeeGFS. ALOMA enhances these tools with domain specific knowledge, such as seismic data patterns and formats, and bundles it with an easy to use workflow generator and execution monitor.