It’s no secret that the upstream oil and gas business is highly capital-intensive. The problem is compounded further by the current inflationary environment where the drilling and labour costs are at multi-year highs. With factors at play, such as the Ukraine-Russia conflict in Europe and years of underinvestment stemming from years of low prices in the 2010s, energy companies are struggling to keep a lid on costs to maximize profits even in an environment of rebounding oil prices. While the effects of climate change are undeniable, the transition to renewable energy sources cannot occur overnight, and fossil fuels will continue to have a purpose. All things considered, organizations are embracing any cost-saving strategy with open arms.
The scalable nature of AI is an excellent use case to reduce the operating cost of oil and gas companies. Once the technology is successfully implemented, the marginal cost is minimal to keep it operational. In 2022, Calgary-based geophysicist Bruce (Nguyen Kieu) published an paper in the journal Improved Oil and Gas Recovery to propose using an AI deep learning model approach to reduce the risk of oil and gas exploration and production by enhancing the resolution of rock property images from seismic forwarding model and inversion processes.
Before oil can be pulled out of the ground, refined, and distributed to gas stations for consumers to fill up their cars, companies must first find the fields and plays where reservoir deposits reside. The “E” part of E&P (exploration & production) is the riskiest part of the business since it is costly to drill an oil and gas well, and the chance of success (positive return on investment) can be very slim. Clearwater in northeastern and central Alberta are some of the most cost-friendly plays in North America where a company may drill a traditional vertical well for roughly $1.5 - $1.7 million. In comparison, multi-stage fracs in the Duvernay or Montney can set a company back $8 million - $ 12 million. The complex logistical nature of offshore drilling makes the drilling cost an even riskier bet, costing upwards of $150 million per well.
Oil and gas companies don’t just poke anywhere on the earth's surface to find oil and gas. The process usually starts with some general geological knowledge of where hydrocarbon deposits can be found, such as the Western Canadian Sedimentary Basin (WCSB) in Western Canada that spans from Northeastern BC to southwestern Manitoba, where many of our proud Calgary companies, such as ARC Resources, Tourmaline, and Canadian Natural Resources operate in. Once a geologic play has been established to contain hydrocarbon deposits, seismic imaging is used in a narrow surface area, where sound is transmitted from the surface to the subsurface, to detect the approximate location of where deposits may be found. It is similar to when Batman in the Dark Knight hacked every cellphone in Gotham to send and receive sonar to build an image of Gotham to locate the Joker. However, the limitation of seismic imaging is that when sound travels from the surface to the subsurface, frequency decays and weakens, and the image built from the sound wave modelling is not perfectly reliable. With modern technology, the success of frontier exploration, a new field thought to contain fossil fuel deposits, is pegged at between 18-30%. Much of the risk and uncertainty lies in the human limitation of interpreting noisy, imperfect seismic imaging. Despite these limitations, seismic imaging is cost-effective as it can cover a large area to approximate the probability of finding energy deposits.
To increase the possibility of exploration success, Bruce proposes using a technique called Smart Seismic Modeling (SSM). The objective is to maximize the value of existing data, enhance seismic resolution, minimize informational noise, and ultimately reduce risk and cost. This technique uses deep learning and combines geophysics, geology, and reservoir engineering expertise to overcome the human limitation in seismic interpretation by enhancing seismic resolution using synthetic seismic forward modeling and artificial intelligence deep learning.
Deep learning attempts to mimic the human brain by feeding input data into successive layers of function nodes, where each node represents an individual regression or classification model. Weights are assigned to determine the significance of a variable with a more considerable weight given to the variable that drives the accuracy of the output. The output of an activation function is used to determine if the result exceeds a specific passing rate (bias). If the output exceeds the passing threshold, the result is fed into the next layer of nodes as an input, which the process is called feedforward. This is the “training” part of machine learning, where inputs turn into outputs and the process iterates through many layers, and the more layers there are in the model, the “deeper” the model gets (hence the name “deep learning”). The ultimate aim is to increase the accuracy and “smartness” of the AI model using an objective cost function by continuously adjusting the weights and biases until the accuracy of a model is maximized and the errors minimized.
SSM has three stages, and the first stage consists of using rock properties data from well logs and the surrounding geological structures to generate a statistically tolerant synthetic seismic dataset. This synthetic dataset is then used as input in a deep learning network to train the model to predict 2D rock physics models using actual seismic gathers and vertical seismic profile (VSP) data for validation. The second stage uses the outputs from stage 1 to prepare a 3D model to make 3D rock-physics model. Stage three is used after a well is in production, and more data is generated from the well to update the models in the first two stages.
To prove the viability of this concept, a simple supervised deep learning (CNN U-Net) model with total of 5 million parameters have been trained over 1,000 epochs on 450 self-generated synthetic training images. Despite of the imperfect matching, the inverted compression velocity (VP) looks positive and prove the potential of using deep learning on rock-physics inversion process. The author acknowledges that there is room for improvements, such as using a more complex convolutional neural network (CNN) or physics-informed deep learning model, using larger dataset with more components (seismic 3C, 4C...), and using more computational power.
As the easy fruits of oil deposits get increasingly picked, companies will need to incur more risk to explore new fields to find more oil. Any way to lower exploration cost should be carefully examined and Smart Seismic Modeling looks promising in the current AI and peak oil age.