SMTI-GPT – A “GPT” for Seismic Moment Tensor Inversion!

SMTI-GPT – A “GPT” for Seismic Moment Tensor Inversion!

Could the technology behind Chat GPT (Generative Pretrained Transformer) be harnessed to solve complex inverse problems in seismology? Absolutely! We have successfully trained a transformer network and integrated it into a broader system to provide automatic and accurate solutions to the full-waveform seismic moment tensor inversion (SMTI) problem.

Our network was trained using synthetic data to reconstruct the six independent components of the moment tensor from normalized samples of the radiation pattern reported to the source. Different modules handle the source time function estimate and Green’s function correction upstream. The source time function is derived from the spectral analysis, and Green’s function correction is applied, considering an arbitrarily complex 3D velocity and attenuation model along the estimated rays. Currently, the site response is not accounted for, as we mainly target mining applications. It should be relatively simple to account for the site’s response in the future.

The moment tensor calculation requires that a

  • P and S velocity models and
  • an attenuation profile or an attenuation function capturing the geometrical spreading

be provided.

The software supports multiple format, howerver, the use of the microseismic data exchange (MDE) format is preferred as information about the system, the event and the waveforms required for the calculation are all integrated into one convenient package. Note that takeoff and azimuth angles at the source are obtained from the rays calculated based on the velocity model.

Using deep neural networks to solve inverse problems is gaining popularity in geophysics. Compared to traditional matrix inversion and optimization methods, deep learning-based approaches offer several advantages:

Solution stability: Deep learning ensures robust and stable solutions, even in the presence of noise and outliers.

Diminished sensitivity to noise and outliers: The network’s capacity to generalize reduces the impact of noisy data.

Computational efficiency and scalability: Deep learning accelerates complex computations, making it ideal for handling large amounts of data.

Ability to capture non-linear relationships: Deep neural networks excel at modelling complex, non-linear geophysical relationships.

Possibility to adapt and refine the model iteratively through transfer learning: Fine-tuning pre-trained models enhances their performance to adapt to specific tasks or to improve the performance through time.

Convenient intrinsic mechanisms to assess solution uncertainty: Deep learning provides inherent mechanisms to estimate uncertainty through ensemble network techniques.

We are now working on deploying the tool in the cloud. An alpha version focusing on automatic solutions with a minimalist user interface will be available shortly for trial.