Toward Fully-Automated In-Field Geophysical Survey Designs using Reinforcement Learning
Abstract
Geophysical data is often acquired using a pre-defined survey design determined using an assumed subsurface model. For surveys requiring the sequential installation of multiple sources and receivers over multiple days in the field, data acquired on one day (and corresponding interim inverted geophysical models) can be used to determine better survey layouts for subsequent days of field operations. However, given the uncertainties and non-uniqueness present in any inverted subsurface model, determining the optimal survey design in such a real-time, in-field manner is a challenging problem. Here we investigate the possibility of online, fully automated survey design for direct current (DC) resistivity surveys using reinforcement learning. Specifically, we use the concept of a Multi-Armed Bandit (MAB), which is a way of finding an optimal action using limited resources within an uncertain environment. The MAB is inspired by the problem a gambler faces at a casino when presented with multiple slot machines ("one-armed bandits") with unknown payout probabilities; what is the optimal way to find the best machine? The DC resistivity survey design problem at hand is formulated by replacing the search for the best slot machine with the search for the best current electrode location, given a fixed array of voltage electrodes; the process is repeated for each new transmitting electrode using the information from all previous measurements to generate an interim subsurface model. We define the "payout" of the survey design using quantities from the data- or model-domain, and examine the numerical behavior of these payouts with and without considering uncertainty in the data and/or assumed model. We then compare results from an MAB optimized layout to conventional survey designs as well as to exhaustive searches of all possible current electrode positions. Results show that this reinforcement learning technique is applicable in the field when such payouts are properly defined.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMNS13A..04N