Artificial Neural Networks (ANNs) are fundamental to modern computational geoscience, particularly adept at analyzing complex datasets, like those obtained from well logs for lithology identification. These ANNs comprise interconnected processing elements, termed nodes or neurons, organized into distinct layers. Each neuron executes specific mathematical operations, and the collective activity across these layers empowers the network to learn intricate relationships within data (LeCun et al., 2015). This architecture allows them to capture and model the nonlinear and intricate relationships found in geological data, making them ideal for predictive tasks where conventional statistical approaches may struggle (Heaton, 2018) (Figure 1).

In the perspective of lithology recognition, a distinctive ANN (Xiong et al., 2020; Xiong et al., 2022; Liu et al., 2024) involves an input layer, numerous hidden layers, and an output layer. The input layer collects raw data like, depth of an interval in meters from the Kelly Bushing elevation, water saturation, density, mass percent bitumen, water resistivity at formation temperature, and volume of shale, which reflect various rock properties. The hidden layers, filled with numerous neurons that possess adjustable weights and biases, process this data. The activation functions like ReLU or sigmoid are used by These neurons to combine non-linearity, serving the network detect complex relationships and interactions indoors the data (Glorot et al., 2011).

Training an ANN implies adjusting the weights and biases to decrease variations between actual and predicted outputs through a method called backpropagation. During this method, the network reduces a predefined loss function (Ng and Jahanbani Ghahfarokhi, 2022) using optimization algorithms like stochastic gradient descent, refining the prediction accuracy (Manouchehrian et al., 2012) with each iteration (Diederik and Jimmy, 2014). Consequently, ANNs have become effective at distinguishing lithological units by detecting subtle differences and characteristics in well log data, serving as an essential resource for geologists and engineers in oil and gas exploration and development.

Featuring Bayesian optimization (Shahriari et al., 2016; Zhang et al., 2020; Asante-Okyere et al., 2022) into ANNs beyond enhances their performance by methodically adjusting hyperparameters (Asante-Okyere et al., 2022; Olmos-de-Aguilera et al., 2023; Nwaila et al., 2024) such as the number of hidden layers, neurons per layer, and learning rates, based on a feedback loop of performance data. This integration not only tailors the network architecture to the unique features of the geological data but also ensures it is optimized for the most accurate lithology predictions, significantly surpassing traditional, non-optimized methods (Snoek et al., 2012).

The acquisition function is used in Bayesian optimization, for global minimum finding on behalf of HF function (hypothesis function) f x . Give leave wherever AF function (acquisition function) f x reaches the highest rate at the present contact x + . The PI function (probability improvement function) is defined as (Shahriari et al., 2016) (Equation 1): P I x = Ρ f x ≥ f x + = ξ z (1)

• z =   μ x − f x + σ x ,

The E I function (EI: expected improvement) can be defined by (Equation 2): E I x = μ x − f x + ξ z + σ x τ z   f o r   σ x > 0 0   f o r   σ x = 0 (2)

• τ : DF (DF: density function) related to normal probability distribution. (Equation 3), (Equation 4) presents: upper confidence bond, lower confidence, bonds:

Bayesian optimization is particularly effective in high-dimensional spaces, which is common in geological datasets where numerous features (F. Xiong et al., 2022) (e.g., well log measurements) must be considered. Traditional methods like grid search are inefficient and often impractical in such contexts, as they require exhaustive evaluations across all combinations of hyperparameters. It uses a probabilistic model to identify areas of the hyperparameter space that are more likely to yield better results. This approach allows it to focus on promising regions rather than blindly exploring the entire space, leading to faster convergence on optimal hyperparameters.

Unlike deterministic optimization methods, Bayesian optimization explicitly accounts for uncertainty in the model predictions. This is critical in lithology identification, where geological data can be noisy and complex. By modeling uncertainty, it allows for more informed decisions about which hyperparameters to test next. The balance between exploration (trying new hyperparameter values) and exploitation (refining known good values) is inherently managed in Bayesian optimization. This is crucial in lithology identification, where understanding the underlying geology often requires exploring various parameter combinations without committing to suboptimal settings.

Geological data often contains noise and outliers. Bayesian optimization’s ability to incorporate uncertainty helps mitigate the impact of such noise on the optimization process (F. Xiong et al., 2022). This resilience allows for more reliable model performance in identifying lithology, particularly in heterogeneous formations. The relationships between input features and lithology classes are often non-linear and complex. Bayesian optimization is well-suited to optimizing models like neural networks that can capture these complexities, allowing for a more nuanced understanding of lithological characteristics, and allows for the customization of the objective function, enabling the incorporation of specific metrics relevant to lithology identification (e.g., accuracy, precision, recall). This customization ensures that the optimization process aligns closely with the project goals, enhancing overall model performance.

In addressing the multifaceted challenge of lithology identification, this study employs Bayesian optimization for neural networks (ANNs) hyperparameters fine-tuning, using its robust probabilistic outline to significantly enhance classification accuracy (Lozano et al., 2011). Well log data encompassing measurements such as depth of an interval in meters from the Kelly Bushing elevation, water saturation, density, mass percent bitumen, water resistivity at formation temperature, and volume of shale are critical inputs for this analysis, reflecting the diverse physical properties of subterranean materials (Jiang et al., 2021). Given the high dimensionality and variability inherent in such data, conventional neural network (Wu and Zhou, 1993) setups without optimization often struggle to achieve optimal performance, underscoring the need for sophisticated tuning methods (Lee et al., 2021).

Bayesian optimization serves as a pivotal advancement in this context, applying a Gaussian process to model the relationship between hyperparameter configurations and their corresponding predictive accuracies (Dutta et al., 2010). This probabilistic approach not only aids in identifying the most effective neural network architecture such as determining the ideal number of layers and neurons per layer but also in fine tuning other critical parameters like learning rates and batch sizes (Snoek et al., 2012). The chosen method utilizes the Expected Improvement (EI) acquisition function, which systematically guides the selection process towards hyperparameter values that are likely to yield improvements over previously tested configurations (Bischl et al., 2023).

Empirical validation of the optimized models on the test dataset discloses that Bayesian-optimized ANNs best traditional, non-optimized counterparts significantly, achieving up to a 96.69% (Validation)/97.21% (Test) accuracy in lithology classification (Asante-Okyere et al., 2022). This score not only underscores the efficacy of the Bayesian approach but also highlights its potential to refine predictive modelling in geoscience applications. Further, statistical analysis, using confusion matrix, confirms the significance of these improvements, reinforcing Bayesian optimization (Zhang et al., 2020) as a crucial tool for enhancing the reliability and accuracy of lithological predictions from well log data (Xie et al., 2023).

The implications of these findings are profound, suggesting that Bayesian optimization (Ng and Jahanbani Ghahfarokhi, 2022) can transform the landscape of geological data analysis by enabling more accurate, efficient, and reliable lithology identification. This advancement promises to reduce the costs and time associated with traditional geological surveys, offering a more streamlined approach that could revolutionize resource exploration and management practices (Nuzzo, 2017; van de Schoot et al., 2021).

The Geological Survey of Alberta started McMurray Formation mapping project, and the Wabiskaw Member overlying of the Clearwater Formation, in the Oil Sands Area of Athabasca. The Alberta Geological Survey data report is one of the most substantial results of the project and will expectantly assist future advancement of the oil sands. The main purpose was to provide a database with 2193 wells data, containing around 750 boreholes that include core analyses, of the Athabasca Oil Sands Area. The current investigation was operated on multiple probing wells (60) that is included in digital data bank that is depicted in a File Reports of Alberta Geological Survey.

The study uses a comprehensive dataset comprised of well log data, sourced from multiple oil and gas fields known for their complex geological settings. This dataset includes several key geophysical logging measurements critical for lithology identification, SitID: A number used within AGS to identify wells, Depth: depth of an interval in meters from the Kelly Bushing elevation, SW: water saturation, PHI: density, W_Tar: mass percent bitumen, RW: water resistivity at formation temperature, and VSH: volume of shale. Each of these measurements provides insights into different rock properties that are indicative of specific lithological characteristics like water resistivity logs that help in identifying fluid content and porosity (Asquith et al., 2004; Ellis and Singer, 2007).

The dataset was carefully curated from a series of drilled wells, each providing a continuous depth-registered record of measured attributes. Preprocessing involved cleaning the data by removing outlier values, interpolating missing data points using statistical techniques, and normalizing the features to a consistent scale to ease effective machine learning analysis (Md Abul Ehsan et al., 2019). The final dataset includes over 18,847 individual log measurements, categorized into several lithological classes based on a combination of core sample analyses and expert geological interpretation. These classes involve Cemented Sand, Coal, Sand, Sandy Shale, Shale and Shaly Sand (Table 1).

The database occurred randomly allotted into training (60%), validation (20%), and test sets (20%). This partitioning ensures that the model is exercised on a representative data sample, validated to tune the model settings without overfitting, and finally tested on unseen data to objectively assess its predictive performance. Such a structured approach to data handling is crucial for developing robust artificial neural network models capable of accurately classifying complex lithological formations (Bishop, 2006).

The optimized and non-optimized artificial neural networks were elaborated for lithology prediction, employing described data on previous sections as inputs with MATLAB R2024a. The inclusive analysis of the ANNs models (2.0: optimized, 3.3: wide, 3.4: bilayer, and 3.5: tri-layer) (Wu and Zhou, 1993) offers a rich illustration of how ANNs architecture and optimization strategies shape the performance of machine learning models (Manouchehrian et al., 2012) in terms of accuracy, cost, and error rates across both validation and test datasets. This detailed exploration reveals the subtleties of model behavior, offering critical insights that can guide the selection of optimal configurations for specific applications, emphasizing the interplay between model complexity, learning strategies, and performance outcomes (Table 2).

Model 2.0 stands out with its custom neural network architecture (Hallam et al., 2022; Ganerød et al., 2023; Neelakantan et al., 2024) and advanced Bayesian optimization (Pavlov et al., 2024). It achieves the highest validation accuracy of 96.69% and maintains substantial effectiveness in the test scenario with an accuracy of 97.21%. Particularly, this model also exhibits the lowest test costs and validation (105 and 499, respectively), as well as the lowest error rates, 3.31% in validation and 2.79% in testing. The greater performance of Model 2.0 can be attached to its advanced optimization technique, which efficiently balances the trade-offs between complexity and performance (Figures 2A, 3A).

Bayesian optimization, acknowledged for its efficiency in exploring parameter spaces, and optimizing performance objectives, improves the model’s capability to generalize across different datasets, thus minimizing overfitting and ensuring robustness. This model reveals how advanced optimization strategies can considerably boost the efficiency of neural networks, making it an ideal choice for applications requiring high accuracy with constrained resource usage.

Model 3.3, described as a Wide Neural Network (Soltanmohammadi and Faroughi, 2023), exhibits somewhat lower performance metrics compared to the other models. With validation and test accuracies of 96.45% and 96.98%, respectively, and higher costs in both scenarios (535 in validation and 114 in testing), it reflects potential limitations in its architectural design and the absence of a specialized optimization approach. The higher error rates of 3.55% in validation and 3.02% in testing further suggest that this model may struggle with efficiency and generalization, potentially due to its wide structure not being complemented by an effective learning strategy. This emphasizes the importance of incorporating advanced optimization techniques for performance improvement of the neural networks, particularly in complex predictive tasks, where precision and cost-efficiency are crucial (Figures 2B, 3B).

Model 3.4, with its Bilayered Neural Network structure, excels particularly in the test dataset, showcasing the highest accuracy of 97.24% and the lowest test cost of 104. This model’s configuration appears to offer an optimal balance, providing sufficient complexity to effectively capture and model intricate data patterns without incurring excessive computational overhead. The low-test error rate of 2.76% underscores its capacity for excellent generalization, suggesting that the bilayered approach is particularly effective in environments where predictive accuracy is paramount, the performance outcomes indicate a potentially well-tuned setup that maximizes efficiency and minimizes costs in operational settings (Figures 2C, 3C).

Model 3.5 represents the most complex network (Ozkaya and Al-Fahmi, 2022; Ommi and Hashemi, 2024) in this analysis, featuring a Trilayered Neural Network. It shows the lowest validation accuracy (96.34%) and the highest validation cost (552), which might indicate a tendency towards overfitting on the validation set due to its deeper network architecture. Nevertheless, it performs well in the test stage, succeeding an accuracy of 97.05% with a relative cost of 111. This suggests that while deeper networks can effectively handle complex datasets, they require careful tuning and possibly more sophisticated optimization strategies to prevent overfitting and manage computational costs effectively. The higher validation error rate (3.66%) further points to the challenges associated with managing more complex models, emphasizing the need for precise model calibration and optimization (Figures 2D, 3D).

In summary, this analysis and comparison delves deep into how different neural network architectures and their associated optimization strategies (Alyaev et al., 2021; Thomas et al., 2023) can dramatically affect machine learning outcomes (Nwaila et al., 2024). The comparison between these models elucidates a spectrum of behaviors and outcomes, from the highly efficient and robust performance of Model 2.0 with its advanced Bayesian optimization to the nuanced challenges faced by the deeper, more complex Model 3.5. Models 2.0 and 3.4 emerge as particularly effective, suggesting that a balanced approach to network design and optimization can yield superior results. This analysis not only grants a described understanding of each model’s strengths and weaknesses, but also extends useful advice for designing neural networks that are tailored to meet specific operational needs and performance criteria (Lawley et al., 2022; Nakamura, 2023). It offers a comprehensive blueprint for leveraging architectural and strategic optimizations to enhance the predictability, efficiency, and cost-effectiveness of neural network models in varied application scenarios.

The application of AI in lithology identification has yielded several key inferences.

Despite the promising findings, several limitations were encountered during the research.

Building upon the insights gained and addressing the identified limitations, several promising future directions for research in AI-driven lithology identification include.

The inferences drawn from recent research in AI-driven lithology identification underscore the transformative potential of AI technologies in revolutionizing geological exploration and resource management practices. Despite encountered limitations related to dataset specificity, computational complexity, and model interpretability, the field continues to evolve with promising advancements in optimization techniques, data integration strategies, and real-time deployment solutions. By adopting these challenges and tracking innovative research paths, the geosciences society can employ the full potential of AI to realize more sustainable, efficient and accurate exploration outcomes in the future.