Accurate and reliable production forecasting is a critical component for the efficient development of oil and gas fields and supports informed scientific decision-making. Although machine learning methods have achieved significant progress in this domain, existing models are typically trained from scratch using limited historical production data, making it difficult to effectively capture the complex nonlinear dynamics, long-term temporal dependencies, and high-dimensional interactions among variables inherent in production time series. This often leads to insufficient generalization capacity and limited predictive robustness. To address these challenges, this study proposes a novel gas well production forecasting method based on large language models (LLMs). The approach builds upon a pre-trained GPT-2 architecture and incorporates several key adaptations to enable effective time-series prediction. First, the input data—including daily gas production rate, tubing pressure, casing pressure, and production time—are subjected to instance normalization to facilitate knowledge transfer. Second, a trainable embedding layer is designed to map numerical time-series data into the semantic embedding space of the LLM, thereby achieving cross-modal alignment between continuous signals and the discrete representation format required by the model. Third, a parameter-efficient transfer learning strategy combining freezing and fine-tuning is implemented: the core self-attention and feed-forward network layers of the LLM are frozen to preserve general-purpose knowledge acquired during pre-training, while the positional encoding and layer normalization modules are selectively fine-tuned to enhance the model’s ability to characterize temporal patterns specific to production dynamics. The resulting model, termed GPT4TS, is systematically evaluated on real-world production data from a marine carbonate gas reservoir in the Sichuan Basin. Experimental results show that for wells with long production histories, GPT4TS significantly outperforms the conventional LSTM model. Under univariate input, the mean absolute percentage error (MAPE) is reduced by 18.573% on average; under multivariate input, the MAPE reduction reaches 35.610%, demonstrating its superior capability in modeling complex trends and leveraging multi-variable synergies. However, for newly commissioned wells with short production histories, insufficient data hinders effective fine-tuning, leading to lower prediction accuracy compared to LSTM. This study not only validates the potential of large language models in petroleum production forecasting but also highlights their strong dependence on historical data length, providing both theoretical insights and practical guidance for model selection in real-world engineering applications.

Traditional production rate forecasting methods are often limited by the production history of individual wells and assumptions of the models, leading to unquantified uncertainties in the prediction results and difficulty in considering the guidance of development patterns from other wells in the block on the target well. Additionally, they fail to fully utilize a large amount of relevant production history data. To address these issues, a new model for probabilistic production rate forecasting driven by multi-well production data is proposed, based on deep autoregressive neural networks. This model integrates dynamic covariate data such as production time and tubing/casing pressure, and employs Bayesian inference along with gradient descent and maximum likelihood estimation methods to derive a generalized historical-future production probability evolution pattern shared among multiple wells. Through data-driven approaches, it achieves probabilistic production forecasting for multiple wells. The performance of the deep autoregressive neural network model is studied using data from 943 tight gas wells in two blocks in the Ordos Basin. Results indicate that compared to traditional deep learning models like LSTM, the new model combines the learned generalized production probability evolution pattern with specific production history data of the target well, forming a “generalized + specific” production probability prediction method. On average, it reduces the relative error by 45% compared to the LSTM model. The classification model reduces the relative error by 24% compared to the global model, further reducing the uncertainty of probability prediction based on the global model and improving the prediction accuracy of specific fine-classified wells. Through validation with actual data, the new model demonstrates better prediction accuracy and stronger robustness, making it applicable for multi-well production forecasting analysis in oil and gas reservoirs.

With the rapid development of deep learning neural networks, new solutions have emerged for addressing fluid flow problems in porous media. Combining data-driven approaches with physical constraints has become a hot research direction, with physics-informed neural networks (PINNs) being the most popular hybrid model. PINNs have gained widespread attention in subsurface fluid flow simulations due to their low computational resource requirements, fast training speeds, strong generalization capabilities, and broad applicability. Despite success in homogeneous settings, standard PINNs face challenges in accurately calculating flux between irregular Eulerian cells with disparate properties and capturing global field influences on local cells. This limits their suitability for heterogeneous reservoirs and the irregular Eulerian grids frequently used in reservoir. To address these challenges, this study proposes a physics-informed graph neural network (PIGNN) model. The PIGNN model treats the entire field as a whole, integrating information from neighboring grids and physical laws into the solution for the target grid, thereby improving the accuracy of solving partial differential equations in heterogeneous and Eulerian irregular grids. The optimized model was applied to pressure field prediction in a spatially heterogeneous reservoir, achieving an average L₂ error and R² score of 6.710 × 10⁻⁴ and 0.998, respectively, which confirms the effectiveness of model. Compared to the conventional PINN model, the average L₂ error was reduced by 76.93%, the average R² score increased by 3.56%. Moreover, evaluating robustness, training the PIGNN model using only 54% and 76% of the original data yielded average relative L₂ error reductions of 58.63% and 56.22%, respectively, compared to the PINN model. These results confirm the superior performance of this approach compared to PINN.

Recent advances in deep learning have expanded new possibilities for fluid flow simulation in petroleum reservoirs. However, the predominant approach in existing research is to train neural networks using high-fidelity numerical simulation data. This presents a significant challenge because the sole source of authentic wellbore production data for training is sparse. In response to this challenge, this work introduces a novel architecture called physics-informed neural network based on domain decomposition (PINN-DD), aiming to effectively utilize the sparse production data of wells for reservoir simulation with large-scale systems. To harness the capabilities of physics-informed neural networks (PINNs) in handling small-scale spatial-temporal domain while addressing the challenges of large-scale systems with sparse labeled data, the computational domain is divided into two distinct sub-domains: the well-containing and the well-free sub-domain. Moreover, the two sub-domains and the interface are rigorously constrained by the governing equations, data matching, and boundary conditions. The accuracy of the proposed method is evaluated on two problems, and its performance is compared against state-of-the-art PINNs through numerical analysis as a benchmark. The results demonstrate the superiority of PINN-DD in handling large-scale reservoir simulation with limited data and show its potential to outperform conventional PINNs in such scenarios.

Artificial neural networks have been widely applied in reservoir engineering. As a powerful tool, it changes the way to find solutions in reservoir simulation profoundly. Deep learning networks exhibit robust learning capabilities, enabling them not only to detect patterns in data, but also uncover underlying physical principles, incorporate prior knowledge of physics, and solve complex partial differential equations. This work presents the latest research advancements in the field of petroleum reservoir engineering, covering three key research directions based on artificial neural networks: data-driven methods, physics driven artificial neural network partial differential equation solver, and data and physics jointly driven methods. In addition, a wide range of neural network architectures are reviewed, including fully connected neural networks, convolutional neural networks, recurrent neural networks, and so on. The basic principles of these methods and their limitations in practical applications are also outlined. The future trends of artificial intelligence methods for oil and gas reservoir development are further discussed. The large language models are the most advanced neural networks so far, it is expected to be applied in reservoir simulation to predict the development performance.

The prediction of gas well performance is crucial for estimating the ultimate recovery rate of natural gas reservoirs. However, physics-based numerical simulation methods require a significant effort to build a robust model, while the decline curve analysis method used in this field is based on certain assumptions, hence its applications are limited due to the strict working conditions. In this work, a deep learning model driven jointly by the decline curve analysis model and production data is proposed for the production performance prediction of gas wells. Due to the time-series characteristics of gas well production data, the long short-term memory neural network is selected to establish the architecture of artificial intelligence. The existing decline curve analysis model is first implicitly incorporated into the training process of the neural network and then used to drive the neural network construction along with the actual gas well production historical data. By applying the proposed innovative model to analyze the conventional and tight gas well performance predictions based on field data, it is demonstrated that the proposed long short-term memory neural network deep learning model driven jointly by the decline curve analysis model and production data can effectively improve the interpretability and predictive ability of the traditional long short-term memory neural network model driven by production data alone. Compared with the data-driven model, the jointly driven model can reduce the mean absolute error by 42.90% and 13.65% for a tight gas well and a carbonate gas well, respectively.

Research activities are currently being conducted to study multiphase flow in hydrate-bearing sediments (HBS). In this study, in view of the assumption that hydrates are evenly distributed in HBS with two major hydrate-growth patterns, i.e., pore filling hydrates (PF hydrates), wall coating hydrates (WC hydrates) and a combination of the two, a theoretical relative permeability model is proposed for gas-water flow through HBS. Besides, in this proposed model, the change in pore structure (e.g., pore radius) of HBS due to effective stress is taken into account. Then, model validation is performed by comparing the predicted results from the derived model with that from the existing model and test data. By setting the value of hydrate saturation to zero, our derived model can be reducible to the existing model, which demonstrates that the existing model is a special case of our model. The results reveal that, under the same saturation, relative permeability to water $K_{rw}$ (or gas $K_{rg}$) in PF hydrates is smaller than that in WC hydrates. Moreover, the morphological characteristics of relative permeability curve (relative permeability versus gas saturation) for WC hydrate and PF hydrate are different.