Deep reinforcement learning for seismic acquisition design in target-oriented imaging

Abstract Seismic acquisition design requires identifying an effective acquisition geometry from a large number of possible configurations. Conventional deterministic optimization methods may become computationally expensive for large search spaces, depend on the initial geometry, and explore only a limited subset of possible configurations. This leads to a combinatorial optimization problem that is generally NP-hard. In this study, we investigate reinforcement learning (RL) as an alternative strategy for target-oriented seismic acquisition design. Under the assumption that the target-reflector image is approximately known, we formulate the source-position design as a Markov decision process and solve it using a deep Q-learning algorithm. For source-position optimization, each state in the RL framework represents a candidate source configuration. Each action modifies the current geometry by moving one source position to the left or right. The reward function is defined by the change in an imaging objective function associated with point-spread functions, thereby encouraging acquisition geometries that improve target illumination and image quality. To make the method computationally feasible, we use Kirchhoff modeling and migration with wave-equation-based amplitudes and traveltimes, and precompute point-spread functions for candidate source positions. Numerical examples show that deep Q-learning identifies source configurations that produce higher-quality target-reflector images than uniform and randomly sampled geometries. These results demonstrate the potential of RL for exploring large and nonlinear seismic acquisition-design spaces.