Wavelet-Activated Quantum Neural Networks Redefine Fluid Typing in Tight Reservoirs

By Aryaman PattnayakReviewed by Susha Cheriyedath, M.Sc.Jan 16 2024

An article published in the journal Scientific Reports explores how accurately identifying subsurface fluids in tight oil and gas reservoirs is a complicated yet crucial process for determining appropriate field development plans and production strategies.

Tight reservoirs, with characteristically low porosity and ultra-low permeability, have become vital unconventional petroleum resources worldwide as conventional sources deplete. However, the complexity of fluid types found in these challenging tight formations, such as dry layers, pure gas layers, water-bearing gas layers with variable saturations, and gas-bearing water layers, creates major fluid typing issues compared to conventional reservoirs.

The typical manual interpretation of well-logging data for fluid identification relies on fallible subjective human judgment. It has inherent disadvantages like low accuracy, low efficiency, and lack of repeatability and auditability. Therefore, developing more intelligent and objective machine-learning approaches for tight reservoir fluid discrimination that can surpass the limitations of qualitative human analysis has become a priority. Various statistical and machine learning methods have previously shown promise in related subsurface problems, including techniques like Bayesian analysis, neural networks, support vector machines, and random forests.

The fledgling field of quantum machine learning, which combines quantum computing principles like superposition and entanglement with deep learning algorithms, is of emerging interest. Quantum neural networks specifically integrate quantum components with the architecture of artificial neural networks, aiming to optimize computational performance beyond classical methods for suitable applications. Research indicates quantum techniques can demonstrate value for industrial evaluation and interpretation tasks on complex multidimensional oil and gas data.

This study focuses on improving fluid typing in tight reservoirs by proposing an enhanced quantum neural network methodology using sensitivity-based input parameter selection alongside wavelet activation functions in the hidden layers. Thorough experimentation compares results to a standard quantum network on field data from China's Sichuan Basin, known for tight oil and gas deposits displaying fluid complexity and nonlinearity unfavorable for conventional identification approaches.

Methods

The core methodology involved analyzing common healthy logging measurements and derived rock physics parameters to determine model inputs with high fluid discrimination sensitivity while assessing correlation to avoid duplicated interrelated information. Calculated sensitivity factors quantified input relevance for each tight reservoir layer type, facilitating the selection of optimized hybrid OK log and rock physics input parameters containing complementary information.

A multilayer feedforward quantum neural network architecture was adopted, with quantum neurons leveraging superposition and interference in the hidden layers. This allowed complex nonlinear fluid behavior to be captured through multidimensional quantum state transitions. Unlike typical quantum neural implementations, wavelet functions were proposed as hidden layer activations, given their superior ability for modeling transient signals over conventional sigmoid functions. Output layer activations remained sigmoidal for bounded final fluid classifications.

Training and testing utilized field data from 70 tight reservoir wells in China's Sichuan Basin, covering over 15,000 footage of various layered fluid types. Inputs for quantum state encoding came from sensitivity-filtered well-logging curves and rock physics properties. The network underwent iterative training with 70% of data, aiming to minimize error against known layer classifications through adjusting interconnect weights and quantum neuron parameters. Testing then evaluated performance on the unseen 30% of tightly held-out well data.

The proposed wavelet-activated quantum network aimed to enhance fluid typing over a regular sigmoid-activated equivalent, with a comparative assessment based on accuracy, precision, recall, and F1 scores over 30 randomized experiments. Additional analysis was conducted on confusion rates between complex coexisting fluid layers like water-bearing gas and gas-bearing water formations. The method offered a novel intelligence-focused solution for an industry-critical tight reservoir analysis task with nonlinear quantum transforms and wavelet localization functions tailored to subtle subsurface signatures.

Results & Discussion

Comparing input configurations showed that hybrid well log and rock physics parameters collectively contained the most information, significantly outperforming individual sets for fluid discrimination when analyzing model training loss profiles. This aligns with the fact that different physics govern various logging responses and rock mechanical properties, complementing each other. The final multifaceted input vector gave versatility in capturing fluid influences.

Testing the trained models over 30 randomized repeats to eliminate sampling bias revealed greater effectiveness overall for the proposed wavelet-activated quantum network compared to the standard sigmoid equivalent. Average accuracy on tight reservoir fluid typing reached 98% using wavelets versus 93% for sigmoids, demonstrating substantially improved performance from the enhanced quantum architecture.

Delving deeper into subclass evaluation, both methods performed excellently for unambiguous dry gas and liquid water layers. However, the wavelet-quantal approach showed improved recall and precision statistics for tricky intermediate water-bearing gas and gas-bearing water formations. This highlights the value of harnessing nonlinear multidimensional wavelet quantum state transitions when classifying ambiguous zones with mixed fluid saturations. Confusion did rise slightly for pure gas layers, likely due to common subtle characteristics between analogous gas-rich formations.

Nevertheless, superior F1 scores substantiate that tight reservoir fluid typing benefits considerably from using the localized time-frequency wavelet functions for activating embedded quantum neural calculations in this problem domain. Reduced confusion and uncertainty intervals translate to better operational decision-making and productivity outcomes when targeting precise layered development.

Future Outlook

This research provides compelling evidence that purpose-built wavelet quantum neural networks can offer a highly accurate and reliable solution for determining complex fluid compositions in log-evaluation tasks on critically important tight subsurface reservoirs.

The techniques show immense promise given that most industry modeling still needs to apply more efficient manual qualitative interpretation, vulnerable to shortcomings like subjectivity, bias, and human fatigue. Adopting intelligent quantum machine learning instead promotes data-driven objectivity, audibility, repeatability, and reduced risk.

While further work remains validating wider deployment across diverse geological settings, these initial Chinese field trial results provide a persuasive argument for the industrial uptake of tuned quantum algorithms. Follow-on enhancement opportunities also exist, exploring incorporating pressure transient tests, special core analysis, and seismic survey data to supplement input parameters.

In summary, as tight oil and gas reservoirs grow increasingly pivotal for global supply security, advanced hybrid quantum techniques like wavelet neural networks demonstrate the enormous capability to leverage information from existing well data accurately and efficiently for optimizing fluid discrimination. Their integration directly facilitates more innovative field development planning, exploration, and drilling target risk reduction in these complex yet critical unconventional reservoirs worldwide.