| Popis: |
Machine Learning has accelerated the ability of seismologists to relate seismic waveform databases to the geological properties of the subsurface through which they travel. With sufficient training, neural networks can infer two and three-dimensional maps of seismic velocity, at a fraction of the computational cost of conventional methods, which are reliant on full waveform forward modeling.Several studies have replaced convolutional neural networks (CNN) with Fourier Neural Operators (FNOs) due to their computational efficiency. In addition to replacing expensive convolution operations with simple multiplication, FNOs take low frequency, low resolution seismic data as input, and still predict successful outputs at the required resolution, provided that the population of waveforms and velocity models used to train them have statistical distributions representative of the ‘ground truth’. One outstanding research question is: how close do seismic datasets need to represent real-world ground truth, to maximise insights from neural networks?This research investigates general approaches to populate training data for FNOs, by creating families of synthetic seismic waveforms in a hybridised statistical-deterministic approach. In practice, this entails accounting for site-specific geological constraints like drilling logs and alternative geophysical surveys. The objective is to improve neural network performance, so that fewer field observations are needed to recover the full seismic model, shifting the collation of massive data to the numerical domain. Here we show results in progress that focus on Irish offshore sedimentary basin settings. |
