Eighth EAGE High Performance Computing Workshop

In computational geophysics, generating precise 3D microstructures from seismic data is crucial for detailed subsurface analysis. Traditional methods often fall short in achieving the necessary resolution, but Generative Adversarial Networks (GANs), specifically SliceGAN, have proven effective.

These networks allow for the generation of large volumes of statistically representative microstructures, improving the simulation of material properties based on their microstructural traits. However, the efficiency of GAN training is critical for both feasibility and accuracy. This study introduces an optimized GAN training method using a distributed data parallel (DDP) strategy within the PyTorch Lightning framework to leverage modern GPU computational power. Significant adaptations to the original SliceGAN code were made to incorporate PyTorch Lightning, allowing for DDP across multiple GPUs, significantly reducing training times and enhancing scalability. The method was tested on NVIDIA V100 and A100 GPUs, demonstrating near-linear scalability and a potential speedup of 48 times with eight A100 GPUs. This optimized training process notably improves the generation of complex, high-fidelity 3D microstructures essential for geophysical analysis, highlighting the advantages of PyTorch Lightning in scenarios requiring high scalability and rapid execution, thereby offering substantial benefits for geophysical research and exploration.

Joint migration inversion, an innovative geophysical method, smartly merges velocity model building with seismic imaging using full wavefield migration algorithms. It uses the full seismic reflection response - including multiple scattering - in a unified framework to produce high resolution geological structure images. Despite precision of FWM, it confronts heavy computational demands, memory needs, and reliance on strong computing resources. This paper proposes a deep learning-based seismic interpolation acceleration strategy centered on a tiny Attention U-Net based model. Designed for efficient frequency domain sparse wavefield interpolation and reconstruction, it reduces full wavefield forward modeling costs. Trained on a 2-D lens-shaped velocity model, the model adaptively learns complex mappings between sparse and complete wavefields, reproducing 2-D numerical simulations while cutting computation time. With up to 50 % missing seismic data, this approach improves efficiency by about 40 % vs. conventional FWM. Integrating trained model into JMI further saves about 30 % compute time under simplified conditions, affirming its advantages and potential in frequency domain forward modeling.

This study utilizes GPU computing to perform seismic full waveform inversion, a high-dimensional and ill-posed problem. By employing hybrid Hamiltonian Monte Carlo methods, it enables efficient computation of the posterior distribution under various priors, which regularize the problem. The approach facilitates assessing the efficiency of different priors and regularization techniques. As high-performance computing continues to advance, these methods allow for the development of more sophisticated inversion algorithms for large-scale seismic problems, improving uncertainty estimation and aiding decision-making.

Devito allows for the generation of optimized GPU kernels. Recently, we managed to code inject Devito generated GPU kernel code in the Shell Wave Equation Library (SWELL) which is used in RTM and FWI. This is done via an interface code, which consists of an interface kernel for the conversion of native SWELL data structures to Devito data structures. The generated Devito kernels are, up to some minor scripting, used as is.

Benchmarking native SWELL kernels against Devito generated kernels gives us a measure of how the native SWELL kernels are performing. Not all complicated propagators can be solved by native SWELL kernels as some of those do not exist. For those propagators it is easier to usegenerated Devito kernels. We started with a 2nd order sponge equation for which we developed the SWELL-Devito interface code and Devito notebooks to generate GPU kernels. The resulting Devito and native SWELL kernels for acoustic constant and VTI constant density propagators are similar in computational speed. Next, the more complicated elastic triclinic propagator (ELTRIVD) was studied. The resulting unoptimized Devito solver for ELTRIVD is around 40/80% slower than the optimized native SWELL counterpart.

This paper outlines the efforts made to integrate HPC profiling tools directly into the user interface of a seismic-processing software package. This endeavor represents a pioneering initiative within the seismic industry, constituting a significant novelty in the processing software landscape. As there is no universal profiling solution capable of offering optimal detail for all scenarios, we meticulously evaluated various profiling tools based on their functionality and usability.

We present an enhanced strategy for multidimensional deconvolution (MDD) that effectively addresses its inherent ill-posed nature. This approach leverages low-rank regularization, which assumes that the unknown Green’s function can be represented with a low-rank structure. By employing lower-rank approximations for lower frequencies and increasing the rank for higher frequencies, our frequency-dependent rank selection achieves a flexible balance between accuracy and memory usage. This adaptive reciprocal low-rank factorization not only improves the accuracy of MDD but also promises significant computational efficiency gains. We demonstrate the effectiveness of our method using a synthetic ocean-bottom cable dataset, paving the way for future applications to large-scale MDD problems.

We focus on accelerating graph processing for seismic data interpretation using GPUs, particularly through the optimization of the Breadth First Search (BFS) algorithm. Seismic interpretation tools like NextVision require processing large graphs, which is traditionally compute-intensive. Initially parallelized on multicore CPUs, the code was slow for large datasets. To improve performance, we implemented GPU acceleration using NVIDIA’s cuGraph library.

The approach involved optimizing the BFS algorithm by launching multiple concurrent searches from independent vertices, maximizing parallelism and reducing overhead. Graph data was efficiently managed using RAPIDS Memory Manager (RMM), and the cuGraph API enabled efficient graph creation and BFS execution without redundant data replication. CUDA multi-streaming was also optimized to improve GPU utilization. Additionally, THRUST API was used to handle dynamic graph updates efficiently.

Experimental results showed significant performance improvements, particularly on larger graphs. The performance platform with an A100 GPU achieved speedups of up to 4.95x for a 44GB graph. Future work will explore additional algorithms, like FastAPSP, to further enhance performance. This study demonstrates the potential of GPU acceleration to significantly speed up graph algorithms in seismic data interpretation, benefiting geophysical analysis.