AI-Powered Software Platforms and Computing Power: Giving BGP the Confidence to Explore "Uncharted Territory"

Petroleum, often referred to as the "lifeblood of the national economy," is an important strategic resource for China.

After decades of extensive exploration, China has located most of its shallow structural reservoirs, leading the country to shift the focus of its oil and gas exploration and development toward deep formations, deep water regions, unconventional resources* , and mature fields. As a result, enhancing petroleum exploration to prove up more reserves has become crucial for safeguarding China's energy "breadbasket."

*Unconventional resources: Shale oil (gas), coalbed methane, and natural gas hydrate

Today, 95% of the world's major oil and gas fields are discovered with seismic exploration technology. This technique involves artificially generating seismic wave to acquire image data of subsurface rock layers. This generated data is then analyzed to identify oil and gas reservoirs, a process often vividly referred to as "giving the Earth a CT scan." The evolution from analog to digital seismic technology, which has been followed by a shift from 2D to 3D seismic imaging, and the recent adoption of wide-azimuth high-density 3D seismic exploration, demonstrates remarkable progress in geophysical exploration technologies. This rapid development has positioned specialized software as a competitive differentiator in oil exploration.

Initiated in 2003 by the Bureau of Geophysical Prospecting INC. (BGP), the subsidiary of China National Petroleum Corporation (CNPC), and refined over two decades of iterative development, GeoEast has emerged as China's first fully-independent, proprietary, and large-scale integrated software system for seismic data processing and interpretation. This marks a historic breakthrough in China's geophysical exploration engineering software, and has been acclaimed as the "Chinese Core" of geophysical technology.

GeoEast leaps forward in both exploration accuracy and efficiency

In seismic exploration, greater data acquisition and more refined processing mean more accurate interpretation of subsurface geology, while significantly improving the chances of discovering oil.

GeoEast supports seismic data processing and interpretation across diverse scenarios, from land to sea, P-waves to multi-waves, and surface to borehole. However, there is a notable side effect: exponential growth in data volume. A single acquisition unit can generate petabytes (PB) of raw data, while individual processing jobs may handle tens of petabytes. Furthermore, seismic data processing involves multiple workflows, frequent I/O operations, computationally-intensive tasks that often take weeks of continuous computation, and a mixture of serial and parallel operations. Consequently, robust storage and high-performance computing capacity are critical requirements that must be addressed.

The GeoEast processing system consists of seven software packages, 18 technical series, and more than 400 functional modules. The source code alone exceeds 30 million lines, making it the largest and most sophisticated piece of industrial software in China. To support for multiple computing architectures in GeoEast and further ensure the security of the GeoEast ecosystem, since October 2023, BGP has partnered with Huawei to launch a computing power adaptation initiative for GeoEast.

Based on the BiSheng compiler, the two teams completed the compilation of 20 million lines of code in just 40 days, and were able to solve compatibility issues in the software. Overall, the entire system was optimized and reconstructed in less than one year.

Since being adapted to Kunpeng, GeoEast has significantly improved its development efficiency and multi-thread computational performance. This is most notable in key high-precision imaging tasks such as pre-stack depth migration and pre-stack time migration, which together account for over 70% of the computational load in seismic data processing. Supported by the Kunpeng-based solution, GeoEast now achieves 2-3 times higher processing efficiency than with the previous processor under comparable operational conditions.

Significant improvements in seismic-wave-solving efficiency accelerate AI adoption in oil and gas exploration

Seismic exploration involves data acquisition, data processing, and data interpretation. Within this process, full waveform inversion* (FWI) is one of the most advanced techniques used for seismic data processing. It achieves the high-resolution reconstruction of subsurface structures by iteratively minimizing the discrepancy between simulated data and observed data. However, traditional FWI is heavily reliant on manually crafted initial models and massive computing power for numerical analysis, making it difficult to apply on a large scale.

*Forward modeling refers to the process of calculating model outputs or predictions based on known model parameters and inputs. Its purpose is to simulate system behavior, predict future states, or validate theoretical models. Inverse problem solving, on the other hand, is the process of deducing model parameters from observed data and a known forward model. Simply put, forward modeling is "from cause to effect," while inverse problem solving is "from effect to cause."

Tradition modeling and imaging

To this end, BGP and Huawei have pioneered the application of AI for Science (AI4S) to FWI, aiming to leverage AI algorithms to address both seismic forward modeling and inversion. This will overcome the computational bottlenecks of traditional FWI and significantly improve its efficiency. Based on the MindSpore AI4S enablement kit, the partnership introduced a dual-data- and mechanism-driven approach for intelligent wave-equation solution. This approach integrates automatic differentiation and diffusion-model-based 3D geological data generation, and adopts a pre-training and fine-tuning workflow to boost solution efficiency and enhance model interpretability and generalization capabilities. Preliminary trial results have shown that a single Ascend accelerator card can complete one iteration of 2D FWI processing for a marine streamer survey line within half an hour. This is a huge improvement over the traditional method, which takes six hours to complete the same task.

To date, verification has been carried out using 2D field data and small-scale 3D model tests, while ongoing efforts are being made to integrate AI4FWI into the GeoEast platform. With AI adoption, computational efficiency regarding 2D and 3D acoustic-wave-equation forward modeling has increased severalfold, and the overall efficiency of FWI has significantly improved, thereby paving the way for industrial-scale deployment of FWI.

Full waveform inversion and imaging

Moving forward, BGP plans to complete the first round of pre-training of a 10-billion-parameter 3D model based on CNPC's Kunlun large model by the end of 2025. BGP is also looking to conduct pilot tests on two or three production projects, as well as validation using both synthetic and field data. These efforts are aimed at accelerating the integration of the AI4FWI module into the GeoEast system and ultimately transforming it into an AI-powered platform.

The industrial adoption of new technologies like AI in oil and gas exploration is set to drive a transformative shift in industry paradigms. Such technologies dramatically reduce turnaround times for seismic data processing and interpretation, while simultaneously enhancing exploration accuracy and efficiency. They also extend our cognitive reach, enabling the discovery of petroleum resources deep underground. Furthermore, they provide new approaches for applying inversion in fields such as geological exploration and medical imaging, thus promoting intelligent transformation across the industry chain.