← Back to Sessions

Methods for Preemptively Optimizing Geophone Array Size for Measurement of Subsurface Volumes using Machine Learning and Synthetic Data From Numerical Simulations

Description: 

Background. Acoustic measurement devices, such as geophones, can be used to image subsurface volumes. After acquiring electro-acoustic seismic data, or shot records, from the geophones, the data must be evaluated using computational methods, such as full-waveform inversion (FWI), to reconstruct the data as a usable image (velocity map). Geophone arrays, which can include thousands of sensors, can be very costly both in capital and operational costs. FWI (particularly in 3D) is computationally expensive and the calculation of residual minima is susceptible to error. Synthetik has developed methods to use machine learning (ML) and physics-based computational methods to identify the minimum hardware requirements to sufficiently image a subsurface volume and circumvent the high computational cost of FWI.

Methods. Synthetik uses a custom numerical solver to create 3D shot records from real or synthetic velocity maps. ML models are trained to perform inversion by using the shot records as input and the velocity maps as ground truths. For an initial pilot study, we evaluated a 39m x 39m x 35m deep volume of variable materials. We simulated 25 equidistant impact shots and a receiver array of 40 x 40 using 1m isometric spacing located at the surface. The mean absolute error (MAE) of the ML model trained and tested using the full recording field (all shots and receivers) was used as a baseline for comparing against field reductions. We then tested decreasing the recording field size by reducing each the sources and receivers fed into the ML model for both training and testing. Abridged Results. Using a trained ML model to solve for a velocity map from full field electrographic data takes >5 seconds each volume with an average MAE of 1.42 for the full field. Reducing the number of receivers in both the x and y planes by a factor of 2 decreased the MAE by 37%. Alternatively, decreasing the source counts by a factor of 5 decreased the MAE by only 35%. Conclusion. Using our methods to identify acceptable levels of accuracy may help reduce measurement and computational requirements, thus reducing imaging time and cost.

Session: Scientific Machine Learning for Forward and Inverse Wave Equation Problems [Poster]

Type: Poster

Date: 4/15/2025

Presentation Time: 08:00 AM (local time)

Presenting Author: David

Student Presenter: No

Invited Presentation: 

Poster Number: 139

Authors