Fast, scalable and economical simulation of three-dimensional time-domain electromagnetic data using survey decomposition and massive parallelization

Abstract Accurate and fast simulation in 3D is critical in quantitatively interpreting time-domain electromagnetic (TEM) data for complex geology, but that task has long been known to be computationally demanding and hence time-consuming, because it involves multiple numerical solutions of 3D Maxwell’s equations at multiple spatial and temporal scales. Survey decomposition (SD) is a generic computational framework that speeds up the electromagnetic (EM) induction problem solutions by breaking down the original large domain, transmitter sources, and time-stepping sequences into many small problems with reduced ranges of scales for the optimal computational efficiency and parallelism. However, the existing SD implementation uses a trial-and-error scheme to determine the number of decomposed sources, creating a sequential dependency that restricts the utilization of ultra-large-scale parallelization for further speed-up. As a result, the promised embarrassing parallelism and linear scalability have not been fully verified and demonstrated. This work first optimizes the large source decomposition strategy by establishing an empirical relationship between the decomposed source, time channels, and conductivity, so the decomposition is complete before the simulation; therefore, the decision-making on the fly can be eliminated. The numerical accuracy of the empirical decomposition is satisfactorily verified against other traditional numerical solutions on a variety of benchmark models. Then, we tested our method on a realistic 3D geological model and a large-loop TEM survey in a parallel computing environment running more than 46,000 cores simultaneously. The linear speedup observed in the numerical experiments demonstrates the excellent scalability of our improved SD method that can fully capitalize the emerging distributed and elastic computing. This study shows that, with the latest algorithms and hardware, 3D TEM modeling problems of any size can now be solved accurately in tens of seconds at a cost of tens of dollars.