An introduction to distributed training of deep neural networks for segmentation tasks with large seismic datasets

Deep learning applications are drastically progressing in seismic processing and interpretation tasks. However, the majority of approaches subsample data volumes and restrict model sizes to minimise computational requirements. Subsampling the data risks losing vital spatio-temporal information which could aid training whilst restricting model sizes can impact model performance, or in some extreme cases, renders more complicated tasks such as segmentation impossible. This paper illustrates how to tackle the two main issues of training of large neural networks: memory limitations and impracticably large training times. Typically, training data is preloaded into memory prior to training, a particular challenge for seismic applications where data is typically four times larger than that used for standard image processing tasks (float32 vs. uint8). Using a microseismic use case, we illustrate how over 750GB of data can be used to train a model by using a data generator approach which only stores in memory the data required for that training batch. Furthermore, efficient training over large models is illustrated through the training of a 7-layer UNet with input data dimensions of 4096X4096. Through a batch-splitting distributed training approach, training times are reduced by a factor of four. The combination of data generators and distributed training removes any necessity of data 1 subsampling or restriction of neural network sizes, offering the opportunity of utilisation of larger networks, higher-resolution input data or moving from 2D to 3D problem spaces.

Bidirectional recurrent neural networks for seismic event detection

Real time, accurate passive seismic event detection is a critical safety measure across a range of monitoring applications from reservoir stability to carbon storage to volcanic tremor detection. The most common detection procedure remains the Short-Term-Average to Long-Term-Average (STA/LTA) trigger despite its common pitfalls of requiring a signal-to-noise ratio greater than one and being highly sensitive to the trigger parameters. Whilst numerous alternatives have been proposed, they often are tailored to a specific monitoring setting and therefore cannot be globally applied, or they are too computationally expensive therefore cannot be run real time. This work introduces a deep learning approach to event detection that is an alternative to the STA/LTA trigger. A bi-directional, long-short-term memory, neural network is trained solely on synthetic traces. Evaluated on synthetic and field data, the neural network approach significantly outperforms the STA/LTA trigger both on the number of correctly detected arrivals as well as on reducing the number of falsely detected events. Its real time applicability is proven with 600 traces processed in real time on a single processing unit.