EAGE Seventh High Performance Computing Workshop

We have developed a novel inverse methodology workflow using Nvidia Modulus’s physics-informed neural operator (PINO) as a fast surrogate to solve the black-oil forward problem together with a modified Bayesian adaptive ensemble Kalman inversion (aREKI) approach parametrized with deep neural network (DNN) exotic priors. The DNN prior used here are the variational convolution autoencoder and a novel mixture of deep neural network experts, and together with the developed aREKI method is optimal for sampling non-Gaussian posterior measures, like situations found in channelized reservoirs. The output from the PINO model is the pressure and water saturation fields, and the inputs are the absolute permeability, effective porosity,time-step,initial pressure & saturation, and source and sink terms. The overall workflow is successful in recovering the unknown channelized absolute permeability & effective porosity field for this synthetic case, using an ensemble of 5,000 members with about 100X speedup to traditional numerical approaches. The workflow is suitable for forward and inverse reservoir uncertainty quantification tasks for rapid reservoir management operational decision strategies.

Numerical simulations are widely used in studying underground processes, as direct measurements are extremely expensive or even unfeasible. Typical examples are the geomechanical processes involved in compacting reservoirs, the evolution of sedimentary basins or the fluid flows in deep, porous formations. Developing efficient, robust and scalable linear solvers to solve those problems is a crucial task. Graphics Processing Units (GPUs) are attracting a growing attention since they are well suited for massively parallel computations providing a very good balance among performance, price and power consumption. In the field of iterative methods, for instance, it is difficult to take advantage from preconditioners based on incomplete factorization and approximate inverses are generally preferred.

In this work, we will focus on the adaptive Factored Sparse Approximate Inverses (aFSAI) that have been already successfully tested on GPUs showing significant speed-ups with respect to the CPU counterpart. We will show in problems arising from geomechanics and basin modelling that, thanks to this GPU- accelerated, it is possible not only to speed-up the simulations but also to significantly reduce the energy consumption to power the hardware.

Modern seismic data acquisition with high-channel count or point-source and point-receiver recording systems can produce massive volumes of data, often exceeding hundreds of terabytes for a single survey. Processing such vast volumes of data becomes challenging due to the limited I/O bandwidth and storage capacity. The use of seismic data compression can help to address these challenges and facilitate more efficient data management. In this paper, we evaluate several cutting-edge lossy compressors, including SZ3, ZFP, and Bitcomp, for compressing seismic data. We compare compression ratios achieved by all compressors using different user-defined error bounds, and explore how compression errors can affect the accuracy of the reconstructed pre-stack gathers and final stacked image. We also examine the difference in compression efficiency between 1D and 2D techniques.

Reverse Time Migration (RTM) is the seismic imaging everyday tool. Traditionally, it is implemented with finite differences method, which is at its core a stencil computation. Stencil computations pose major challenges for all major processor architectures with their deep memory hierarchies. In this work, a novel RTM algorithm is presented. It efficiently exploits extreme scale distributed-memory platforms with no cache hierarchy, a paradigm closer to dataflow rather than traditional von Neumann architecture. The implementation following this paradigm delivers disruptive level performance.

Seismic imaging and interpretation are important workflows for the Oil and Gas (O&G) industry, offering critical subsurface insights for informed decision-making. This paper presents a case study of SEAM Subsalt TTI Model that effectively harnesses the advanced capabilities of HPC, cloud platforms, and AI/ML aided interpretation to improve and complement the traditional O&G industry workflow processes.

We have tested our pilot 3D generative modeling and hybrid physics + AI/ML mapping applied to oil & gas reservoir navigation. Initially, we developed the capability to statistically mimic a logging survey and analyze the computed electromagnetic tool responses. Then, under real-time and pre-/post-well memory workflow conditions, probabilistic inverse mapping has been investigated aiming at optimal placement of the new wells. We wrap-up our forward and inverse applications into deployable soft notebooks and power them with CPU/GPU parallel computing and 3D scene visualization. The use of supercomputing-on-chip performance features, web-based GUIs and OS-agnostic environments makes these soft notebooks very suitable for fast-fail-fast-learn interactions not only with internal subject-matter experts, but also with early adopters among the partners and customers in the industry.

Despite its simplicity, the gradient reduction stage can become a bottleneck in the execution of FWI (and RTM as well). Leveraging the file system for this stage can lead to very simple and easy to develop architectures, but performance can be limited and variability of performance can be significant. Leveraging the network may increase perfromance and reduce varibility, but the complexity to be managed is relevant. In this abstract, the optimization of this stage is analyzed and discussed

With the advent of large offset and low frequency seismic data, the information stored in surveys has become ever richer and more voluminous. At the same time, a push for more detailed solutions requires the inclusion of higher frequencies from the data. Moreover, to support extracting accurate and realistic geophysical models of the subsurface, velocity model building such as done in Full Waveform Inversion (FWI) frequently requires inclusion of anisotropic parameters, elastic, and viscous information. However, the computational cost associated with solving such realistic equations is non-trivial.

In the current work, we have implemented the fully anisotropic viscoelastic equations (and all simpler cases). We use a velocity-stress staggered grid approach proposed in [ 5 ] with optimized FD weights of any spatial order and second order in time. We discuss performance for multi-GPUs & multi-node GPUs, as well as some of the optimization techniques used for convolutional perfectly matched layer (CPML) and MPI.

We propose a runtime method to partial or fully override the MPI inside the container, by one version that is optimized for the target machine. Our approach does not require a container image rebuild/update and doesn’t require a match between the host and the container OS. We executed a high order 3D stencil using two nodes with two MPI processes per node (PPN) to demonstrate the performance difference by the original container with Intel MPI, and an overridden container with Cray and MVAPICH2 tuned for the target machine Slingshot fabric.