Research On Geophysical Inversion Based On Reinforcement Learning

With the increasing demand for national economic construction and development for mineral resources,the existing geophysics data inversion and interpretation methods have fallen behind the actual exploration demand.Reinforcement learning is good at transforming abstract combinatorial optimization problems into Markov Decision Process in an environment-driven way,and using concepts such as State,Action,and Value to mathematically and model the solution and optimization process so that the combinatorial optimization problem,which was initially impossible to start from,has a complete set of mathematical theories.Combinatorial optimization is a kind of optimal subset problem that seeks to achieve a specific objective in discrete space,while the goal of geophysical inversion is to find a set of geophysical parameters of all underground grid blocks to minimize the fitting difference between the forward results of the inversion model and the observed data.In this case,geophysical inversion is also a combinatorial optimization problem in nature.It has a pioneering and universal value to explore the application of reinforcement learning in geophysical inversion.This paper will systematically introduce the technical details of Q-Learning algorithm and DQN algorithm to solve the problem of 3D magnetic inversion.The main research achievements and contributions are as follows:(1)This paper offers a set of algorithm frameworks for 3D magnetic classification inversion based on the reinforcement learning Q-learning algorithm.Besides,this paper designs a series of experiments,such as the single prism model,double prisms model,single inclined plate model,fault model,and fault model with shallow noise interference,to verify the reliability and inversion effect of the algorithm.In addition,this paper also proposes an index to describe the algorithm's confidence--profile probability graph,which provides a new perspective for investigating the algorithm's inversion effect.(2)Based on the calculation formula of vertical cuboid magnetic field proposed by Shrama,this paper re-deduces and proposes the tensor form of vertical cuboid magnetic anomaly three components,which is more suitable for programming and parallel computing.This method does well in the forward calculation of complex model magnetic field formed by a large number of vertical cuboids stacked together.Besides,as the training process involves lots of magnetic anomaly calculation,this method could speed up the whole inversion process.(3)Based on the Deep Reinforcement Learning DQN algorithm,this paper proposes a set of algorithm framework for fuzzy edge correction in Q-learning algorithm inversion results,and this method is verified by the fault model.