The integration of AI with geosciences narrates a story of dynamic convergence between technology and specialized knowledge. At the dawn of this collaboration, early systems operating on predefined rules entered the realm of geological data interpretation. Although these pioneering systems faced numerous limitations, they opened new horizons for using computational intelligence to analyze the most complex geological problems [6].

Moving into the 1990s, a new generation of intelligent systems based on Artificial Neural Networks (ANNs) emerged. These computational architectures, inspired by the human brain structure, demonstrated remarkable ability in identifying complex patterns within well log data and opened a window toward modeling subsurface heterogeneities [1], [4], [6].

On the threshold of the third millennium, the third generation of intelligent technologies, including Support Vector Machines (SVMs) and ensemble learning methods, swept across the field. These advanced architectures, utilizing complex mathematical concepts, demonstrated unprecedented capability in analyzing multidimensional data and identifying nonlinear relationships [5].

In recent years, the DL revolution has also permeated geosciences. Convolutional Neural Networks, inspired by the human visual cortex, have gained the ability to intelligently extract meaningful features from seismic data, while advanced recurrent architectures have enabled complex temporal analyses [7].

Today, the newly emerging generation of intelligent systems follows an integrated approach combining advanced data mining with fundamental physical laws. These fourth-generation systems not only learn from data but also incorporate the fundamental concepts governing geological phenomena within their architecture [8].

Furthermore, the necessity for transparency in AI decision-making has led to the development of interpretable technologies. These technologies enable geoscience specialists to understand the logic behind intelligent system predictions and utilize them with greater confidence [9].

The future outlook envisions the development of distributed intelligent systems that can learn collaboratively while maintaining data privacy. This transformation, along with the automation of intelligent model development processes, will create a revolution in the speed and quality of geological analyses.

This dynamic evolutionary path has not only transformed the accuracy and efficiency of geological analyses but has also provided new foundations for understanding Earth's most complex mysteries [10].

The process of lithology prediction has undergone a significant transformation over time. Initially, geologists used simple mathematical relationships and empirical charts for lithology identification. While these methods were understandable, they faced limitations when dealing with complex geological formations.

With technological advancements, methods based on classical statistics, such as discriminant analysis and linear regression, have replaced traditional approaches. These techniques enabled quantitative classification of rocks but were unable to model nonlinear relationships present in well log data [10–12].

Subsequently, decision tree algorithms and SVMs emerged. These methods, with their capability to identify more complex patterns, enhanced prediction accuracy. However, they still had limitations in extracting relevant features from multidimensional data.

The introduction of ANNs marked a turning point in this field. These networks demonstrated greater flexibility in lithology prediction through their ability to learn complex nonlinear relationships. However, the real progress came with the emergence of DL architectures [2], [4–6].

Convolutional Neural Networks, with their capability to automatically extract spatial features from well log data, created a fundamental transformation in identifying lithological patterns. Simultaneously, Recurrent Neural Networks, with their ability to analyze time sequences, enabled modeling of vertical variations in rocks [9].

In recent years, combining these architectures in the form of hybrid models has yielded significant results. Furthermore, integrating well log data with seismic information within multimodal learning frameworks has provided a more comprehensive understanding of lithological systems.

The development of interpretable methods alongside improved model accuracy has enabled verification and validation of results by expert geologists. These advances have increased trust in AI predictions and expanded their practical application in exploration projects.

Today, research focuses on developing models that can train with limited data and incorporate existing geological knowledge into the learning process. These approaches pave the way for AI applications in data-scarce regions [9], [11], [12].

The estimation of shear wave velocity (Vs) has progressed through several distinct phases, each marked by increasing methodological sophistication. Initially, the industry relied heavily on simplistic empirical relationships that linked Vs directly to compressional wave velocity (Vp) using fixed coefficients. These early models, while straightforward to apply, often failed to account for the complex lithological and petrophysical variations inherent in subsurface formations, leading to significant inaccuracies in heterogeneous reservoirs [1], [6], [13].

The introduction of nonlinear regression techniques represented a substantial leap forward, enabling the capture of more complex interactions between petrophysical parameters. However, the true transformation began with the adoption of ANNs, which could learn intricate patterns from well log data without relying on predetermined mathematical forms. Early ANNs demonstrated remarkable improvements in prediction accuracy but were limited by their susceptibility to local minima and requirements for large training datasets [5].

Contemporary approaches have embraced more advanced DL architectures. Deep Belief Networks (DBNs) and Restricted Boltzmann Machines (RBMs) have shown exceptional capability in capturing hierarchical features from complex petrophysical data. These models excel in identifying subtle patterns that traditional methods overlook, particularly in challenging environments like organic-rich shales and complex carbonate reservoirs. Furthermore, the integration of physics-based constraints with data-driven approaches has emerged as a powerful paradigm, ensuring that predictions not only match the data but also comply with fundamental rock physics principles [1].

Current research focuses on hybrid models that combine the strengths of multiple algorithms. For instance, integrating convolutional neural networks for spatial feature extraction with recurrent networks for temporal dependencies has yielded significant improvements in prediction accuracy across varying geological contexts. The incorporation of transfer learning techniques has also enabled effective model application in data-scarce environments, addressing one of the persistent challenges in shear wave velocity estimation [1], [6], [13].

The journey of TOC prediction methodologies reflects a continuous pursuit of greater accuracy and efficiency. The earliest approaches depended entirely on direct laboratory measurements through pyrolysis and chemical analysis. While providing fundamental ground truth data, these methods were constrained by high costs, time requirements, and limited spatial coverage, making comprehensive reservoir characterization challenging.

The development of empirical relationships based on petrophysical logs marked a significant advancement, allowing for continuous TOC estimation along wellbores. Log-based parameters such as gamma ray, resistivity, and density were combined through mathematical formulations to infer organic richness. However, these methods often struggled with geological complexity and required extensive calibration specific to each basin or formation.

The advent of AI revolutionized TOC prediction by introducing nonlinear modeling capabilities. Early ANNs and fuzzy systems demonstrated the ability to handle complex, non-linear relationships between multiple well logs and TOC measurements. These systems could adapt to varying geological conditions and provided substantially improved predictions compared to traditional empirical methods.

Modern TOC prediction leverages sophisticated ensemble methods and DL architectures that integrate diverse data types. Current best practices involve combining conventional well logs with advanced spectroscopy data, seismic attributes, and geological constraints through multi-modal learning frameworks. These integrated approaches have proven particularly [2], [5], [13].

Emergence and evolution of AutoML. The concept of AutoML first emerged in response to the growing complexity of AI modeling and the need to make AI accessible to non-specialists. In initial stages, systems like Auto-WEKA and Hyperopt focused primarily on automating parameter configuration for classical algorithms such as SVM and Random Forests (RF). These early systems employed advanced techniques like Bayesian Optimization (BO) and evolutionary algorithms to navigate complex parameter spaces.

Over time, AutoML frameworks evolved toward comprehensive automation of the ML pipeline. Examples like TPOT utilized genetic programming to optimize complete modeling pipelines, while user-friendly platforms such as Google Cloud AutoML enabled non-experts to leverage advanced AI capabilities. During this period, research focus shifted toward Neural Architecture Search (NAS), with methods like ENAS and DARTS making efficient neural network architecture design feasible through reduced computational costs. The subsequent evolution of the field came with the emergence of AutoDL, which enabled automatic design of deep neural networks for specialized applications. Frameworks such as Auto-Keras and H2O.ai represented successful implementations of this approach. Recent research, such as the study on the World Hyper-Heuristic algorithm, has further advanced AutoML capabilities by integrating reinforcement learning and meta-heuristic algorithms.

In geosciences, while AutoML applications remain in early stages, they have shown promising results in areas like lithology classification and reservoir modeling. Notable achievements include reduced dependency on manual feature engineering and improved model accuracy. However, fundamental challenges such as managing heterogeneous geological data and integrating domain knowledge into the automation process still require innovative solutions.

Future horizons for AutoML in geosciences include developing multi-objective systems to balance accuracy, interpretability, and efficiency, as well as designing federated architectures for processing distributed data. By leveraging capabilities of advanced algorithms like reinforcement learning and hyper-heuristics, AutoML has the potential to become the core component of next-generation geoscientific research workflows [1].

Hyper-Heuristic algorithms have fundamentally transformed computational optimization approaches. These methods have passed through three main generations and have today reached significant maturity.

Generation 1 (Simple selection architectures): In the initial phase, these algorithms operated based on elementary selection mechanisms. The core of these systems relied on conditional rules and random patterns to manage heuristic libraries.

Generation 2 (Learning systems): The emergence of this generation coincided with the integration of intelligent computational methods. During this period, the capability to generate new heuristics and dynamically adapt to problem characteristics was added to the systems. Additionally, knowledge transfer between different problem domains was made possible.

Generation 3 (Integrated intelligent architectures): The World algorithm, as a representative of this generation, offers advanced capabilities:

I. Intelligent analysis of the search space and selection of the optimal strategy.

II. Dynamic adjustment of exploration and exploitation parameters.

III. Experimental learning from previous optimization processes.

IV. Integrated processing of discrete and continuous problems [3].

The rapid integration of modern digital technologies has fundamentally reshaped the methodologies and scope of geoscientific research. This transformation is primarily driven by the synergistic interplay of AI, Cloud Computing, and the Internet of Things (IoT), collectively forming a transformative technological triangle that enhances the precision, scalability, and intelligence of earth-science investigations.

Through this integration, real-time monitoring and adaptive reservoir management have become feasible, enabling continuous data acquisition and informed decision-making. Furthermore, predictive analytics applied to exploration and production systems facilitates proactive maintenance and optimization of resource extraction processes. AI-based algorithms also contribute to early detection of operational anomalies, significantly improving safety and efficiency across geoscientific operations.

Another frontier technology, Quantum computing, has demonstrated remarkable potential in addressing the computational challenges inherent in complex geological modeling and inverse problem-solving. Early applications suggest that quantum algorithms can substantially accelerate high-dimensional simulations and parameter estimations that were previously constrained by classical computation.

Complementing these advances, Digital Twin frameworks now enable the accurate virtual replication of geological systems by integrating physical models with real-time field data. This coupling supports dynamic simulation, scenario analysis, and predictive forecasting, fostering a deeper understanding of subsurface dynamics.

Collectively, these developments establish a robust digital infrastructure that underpins a new era of geoscientific discovery—one characterized by comprehensive, data-driven, and high-fidelity solutions to the multifaceted problems of earth sciences [1], [4], [6], [13].

Although recent advances in AI have brought about remarkable transformations in the geosciences, numerous challenges still remain for researchers in this field. One of the most fundamental of these challenges is the issue of the quality and quantity of available data. Well log data are often incomplete, noisy, and heterogeneous, which significantly impacts the accuracy of AI models. Furthermore, the computational complexity of advanced models, especially when processing massive three-dimensional data, requires powerful and expensive hardware resources, which itself poses a barrier to the widespread development and deployment of these technologies.

Another challenge is the interpretability of results. Complex DL models often function as "black boxes," making it difficult for geologists to understand their decision-making logic. This lack of transparency reduces specialists' trust in AI predictions. Additionally, integrating heterogeneous data from diverse sources, such as seismic data, well logs, and core samples, presents further problems due to differences in scale and resolution. Transferring knowledge from laboratory settings to actual field conditions also requires more research, as environmental factors like temperature and pressure can significantly alter rock behavior [1].

In the face of these challenges, future trends appear promising. It is anticipated that in the near future, we will witness the development of more automated algorithms that minimize the need for human intervention. AI systems will be capable of fully automating the model design process, from architecture selection to parameter tuning. The convergence of AI with other emerging technologies, such as the IoT and cloud computing, will also create profound transformations in the geosciences.

Intelligent sensor networks that collect field data in real time, coupled with cloud systems offering unprecedented processing power, will enable smarter monitoring and management of subsurface resources. The emergence of more advanced technologies (like quantum computing) could make complex geological optimization problems, previously unsolvable, tractable. Furthermore, the development of transfer learning and multi-task learning methods will allow AI models to be trained with less data and higher efficiency. Finally, there is a growing focus on developing sustainable and environmentally friendly solutions. This trend includes optimizing the energy consumption of algorithms and using AI for the sustainable management of natural resources. Collectively, these developments paint a future where AI will serve as an intelligent and reliable collaborator in advancing the geosciences.

Looking at the history of AI in geosciences, a clear evolutionary path can be observed. From simple expert systems to advanced hyper-heuristic algorithms, this field has constantly been developing. The current research, continuing this evolutionary path, attempts to solve complex geoscience challenges by employing the advanced World algorithm.

The proposed methodology integrates the World Hyper-Heuristic algorithm with AutoDL to optimize neural network architectures for geological modeling tasks. The central objective is to achieve automated, adaptive, and generalizable optimization of neural networks for lithology classification, shear wave velocity (Vs) prediction, and TOC estimation.

The World algorithm operates as a reinforcement-driven global optimization system, inspired by ecological and social interactions within a global population. Each “world” represents a micro-environment containing multiple solution agents that interact, compete, and cooperate according to defined rules of energy exchange and knowledge transfer. The system dynamically adjusts its search behavior by balancing exploration (diversity-oriented search) and exploitation (intensified local optimization) to prevent premature convergence and enhance solution robustness across non-convex, multimodal landscapes typical of geoscientific data. This methodological framework is designed to:

I. Automate neural network architecture search and hyperparameter tuning within Auto-Keras.

II. Integrate reinforcement learning–based control for adaptive optimization.

III. Employ geological domain constraints to maintain physical plausibility and interpretability of results.

The World Algorithm is built upon four core mechanisms that emulate natural global dynamics:

I. Population initialization: a diverse set of candidate solutions (agents) is randomly generated, each representing a neural network configuration (number of layers, activation functions, regularization parameters, etc.). II. Global interaction dynamics: agents interact across multiple “worlds.” Within each world, local cooperation enhances exploitation; inter-world communication introduces new genetic material to preserve diversity. III. Energy-based evaluation: each agent’s energy corresponds to its fitness, defined as a weighted combination of model accuracy, loss, and computational efficiency.

IV. Adaptive learning and migration: the system employs reinforcement signals derived from performance improvements. High-performing agents broadcast their structural patterns to other worlds, while weaker agents undergo transformation (mutation or migration).

Mathematically, the optimization objective F(θ) is expressed as a multi-objective function:

( ) = 1 ⋅ ( ) − 2 ⋅ ( ) − 3 ⋅ ( ), where Acc(θ) represents predictive accuracy, Loss(θ) the training error, and CompCost(θ) the normalized computational overhead. The adaptive weights wi are dynamically updated based on reinforcement learning feedback to prioritize objectives depending on optimization progress.

To ensure full automation, the World algorithm is integrated within the Auto-Keras framework as a global controller for both NAS and hyperparameter tuning. The search space includes:

I. Number of hidden layers (1–10).

II. Neuron counts per layer (16–1024).

III. Activation functions (ReLU, Tanh, ELU, LeakyReLU).

IV. Optimizers (Adam, RMSProp, SGD, Nadam).

V. Regularization schemes (Dropout 0.1–0.5, L2 penalties).

VI. Learning rates (10⁻⁵–10⁻²).

Auto-Keras executes the base learning process, while the World algorithm adaptively navigates this configuration space using feedback loops that incorporate both accuracy and computational cost metrics.

Two distinct datasets were utilized to evaluate the method:

I. Dataset A: North Sea reservoir wells, comprising gamma-ray, resistivity, density, neutron porosity, and sonic logs with lithology labels.

II. Dataset B: Horn River Basin shale dataset containing multi-well measurements with corresponding shear wave velocities and TOC laboratory values.

Each dataset was subjected to standardized preprocessing:

I. Missing values were imputed using median interpolation.

II. Outliers were removed using an interquartile range filter.

III. Features were normalized to zero mean and unit variance.

IV. Data were split into training (70%), validation (15%), and test (15%) subsets, maintaining stratified sampling to preserve lithological proportions.

Model performance was assessed using multiple quantitative metrics to ensure comprehensive evaluation:

+ + +

= ∑ ( − ̂ ) ,

∑( − ̂ ) = 1 − , ∑( − ˉ)

Additionally, computational efficiency (measured as training time and FLOPs) and stability across multiple runs were recorded to evaluate the robustness of the optimization process.

To validate the proposed method, the World algorithm’s performance was compared against several established optimizers:

I. Genetic Algorithm (GA).

II. Particle Swarm Optimization (PSO).

III. BO.

IV. Tree-Structured Parzen Estimator (TPE).

Each optimizer was implemented under identical experimental settings to ensure fair comparison. Statistical significance of results was confirmed using the Wilcoxon signed-rank test (p < 0.05).

All experiments were conducted on a cloud-based environment equipped with NVIDIA Tesla A100 GPUs, 128 GB RAM, and TensorFlow 2.15 backend. The World algorithm was implemented in Python 3.10, integrated through Auto-Keras’ custom tuner API. A total of 200 optimization iterations were executed per dataset, with each iteration comprising 30 parallel evaluations.

This methodology advances the field of geoscientific AI by:

I. Introducing a global interaction–inspired metaheuristic tailored for neural architecture optimization.

II. Demonstrating reinforcement-driven adaptability that enhances search efficiency in complex geological landscapes.

III. Enabling fully automated model design through seamless integration with AutoDL frameworks.

IV. Providing a generalizable and interpretable optimization framework applicable across multiple geological prediction tasks.

This section presents the experimental outcomes of applying the World algorithm to optimize neural network architectures across two geological prediction tasks: 1) lithology classification, and 2) shear wave velocity (Vs) and TOC prediction. The results demonstrate that the proposed algorithm outperforms conventional metaheuristics in terms of accuracy, convergence speed, robustness, and computational efficiency.

The World-optimized network achieved the highest performance across all evaluation metrics. As summarized in Table 1, it reached an average accuracy of 96.3%, surpassing PSO (91.5%), GA (90.7%), BO (93.2%), and TPE (92.6%).

Compared with standard metaheuristics, the World algorithm demonstrated a ~5% improvement in accuracy and ~30–40% faster convergence. The reinforcement-driven balancing between exploration and exploitation enabled the model to avoid stagnation in local optima and to adaptively refine the search direction based on agent-level rewards.

The confusion matrix (Fig. 1) shows that the World-tuned model achieved near-perfect discrimination among sandstone, shale, and carbonate classes, with minimal confusion between lithologically similar formations.

For regression-based geological properties, the World algorithm also outperformed all benchmark optimizers in predicting continuous targets such as Vs and TOC.

The World-optimized network yielded R² values above 0.96 for Vs and 0.94 for TOC, marking a substantial improvement over the closest competitor (TPE). The decrease in RMSE by ~15% confirms that the model not only fits the data more precisely but also generalizes better to unseen samples.

The enhanced generalization capability stems from the algorithm’s dynamic energy exchange and inter-world cooperation, which maintain population diversity while focusing computational resources on promising regions of the search space.

Fig. 2 illustrates the convergence trajectories of all optimizers averaged over five runs. While GA and PSO exhibited oscillatory convergence due to over-exploration, Bayesian and TPE methods tended toward early stagnation. In contrast, the World algorithm demonstrated a smooth yet rapid convergence curve, stabilizing after approximately 110 iterations compared to 150–170 for the others.

Furthermore, the standard deviation of final fitness across runs (σ = 0.0048) was significantly lower than that of PSO (σ = 0.012) and GA (σ = 0.015), confirming its superior stability and reproducibility.

Beyond accuracy, computational efficiency is critical for real-world deployment. The World algorithm reduced total optimization time by nearly 30% compared to GA and PSO. This efficiency gain is attributed to:

I. Reinforcement-guided pruning of low-potential agents.

II. Parallel multi-world interactions that distribute exploration efforts effectively.

III. Adaptive migration policies minimizing redundant evaluations.

Profiling analysis revealed that despite maintaining a larger population diversity, the algorithm’s intelligent task scheduling and feedback-based pruning kept GPU utilization high (87%) and computational waste low.

To ensure statistical significance, all metrics were subjected to the Wilcoxon signed-rank test (α = 0.05). The World algorithm’s superiority was found statistically significant over GA, PSO, and BO (p < 0.01), and marginally over TPE (p = 0.047).

The effect size (Cohen’s d) of performance gain ranged between 0.82 and 1.04, indicating a strong to very strong improvement magnitude.

To assess interpretability, SHapley Additive exPlanations (SHAP) analysis was conducted on the final trained models. The most influential features for lithology classification were gamma-ray and density logs, while Vs prediction was primarily driven by compressional velocity and resistivity.

The feature importance ranking remained consistent across runs, verifying that the World algorithm’s optimization process does not compromise interpretability, a crucial advantage for geological decision-making.

The results collectively demonstrate that the World algorithm consistently achieves superior accuracy, stability, and computational efficiency compared with traditional metaheuristics. Its reinforcement-driven global cooperation enables balanced exploration and exploitation, making it highly suitable for high-dimensional, non-convex geological modeling problems.

In summary:

I. Accuracy improved by 3–6% on classification tasks.

II. RMSE reduced by 12–18% on regression tasks.

III. Convergence accelerated by 30–40%.

IV. Variance across runs dropped by more than 60%.

These findings substantiate the World algorithm as a state-of-the-art global metaheuristic for optimizing deep neural models in geoscience applications.