Wenyue Sun
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Wenyue Sun has a total of 64 co-authors. Here are the top co-authors by number of shared publications:
- Sathish Sankaran
Xecta Digital Labs Ltd
4 shared publications - Shuyang Liu
Ministry of Education
4 shared publications - Hangyu Li
School of Petroleum Engineering, China University of Petroleum (East China), Qingdao 266580, China
3 shared publications - Ming Gao
School of Petroleum Engineering, China University of Petroleum (East China); Key Laboratory of Unconventional Oil & Gas Development, China University of Petroleum (East China), Ministry of Education
3 shared publications - Bin Gong
China University of Geosciences, Wuhan, China
2 shared publications - Hoonyoung Jeong
Department of Energy Resources Engineering College of Engineering Seoul National University Seoul Korea
2 shared publications
Latest Publications
- Open Access
Reservoir Modeling Assisted with Knowledge-Enhanced Large Language Models
Reservoir simulation is an important technique to characterize subsurface fluid flow and to support reservoir development and production optimization. Conventional reservoir modeling requires engineers to extract model-relevant information from heterogeneous data, organize the information into predefined formats, and generate simulator-specific input files through professional software. Constructing and iteratively updating such models can often take months of effort due to the high modeling complexity. Large language models (LLMs) offer a potential alternative by translating modeling requirements into simulator input instructions through natural language modeling instructions. However, proprietary reservoir simulators are poorly represented in public training corpora, and their keyword systems may differ substantially from those of widely documented commercial simulators. This study addresses these challenges using the Compositional Advanced Simulation System (COMPASS), an in-house reservoir simulator for unconventional resources, as the target simulator. The proposed knowledge-enhanced LLM framework combines retrieval-augmented generation (RAG) with schema-constrained tool calling to support well-control instruction generation. On the 50-task RECURRENT benchmark using DeepSeek-V4.1-Flash, the four generation strategies achieved average accuracies of 0.67%, 74%, 98%, and 100% for Strategy I (zero-shot), Strategy II (RAG), Strategy III (schema-constrained tool calling), and Strategy IV (RAG combined with schema-constrained tool calling), respectively. Additional experiments with different base LLMs showed accuracies ranging from 92.0% to 100.0%. These results demonstrate that combining enterprise-domain knowledge retrieval with schema-constrained output generation can substantially improve the reliability of LLM-based generation for proprietary reservoir simulation software.
Open AccessReservoir Modeling Assisted with Knowledge-Enhanced Large Language Models
Reservoir simulation is an important technique to characterize subsurface fluid flow and to support reservoir development and production optimization. Conventional reservoir modeling requires engineers to extract model-relevant information from heterogeneous data, organize the information into predefined formats, and generate simulator-specific input files through professional software. Constructing and iteratively updating such models can often take months of effort due to the high modeling complexity. Large language models (LLMs) offer a potential alternative by translating modeling requirements into simulator input instructions through natural language modeling instructions. However, proprietary reservoir simulators are poorly represented in public training corpora, and their keyword systems may differ substantially from those of widely documented commercial simulators. This study addresses these challenges using the Compositional Advanced Simulation System (COMPASS), an in-house reservoir simulator for unconventional resources, as the target simulator. The proposed knowledge-enhanced LLM framework combines retrieval-augmented generation (RAG) with schema-constrained tool calling to support well-control instruction generation. On the 50-task RECURRENT benchmark using DeepSeek-V4.1-Flash, the four generation strategies achieved average accuracies of 0.67%, 74%, 98%, and 100% for Strategy I (zero-shot), Strategy II (RAG), Strategy III (schema-constrained tool calling), and Strategy IV (RAG combined with schema-constrained tool calling), respectively. Additional experiments with different base LLMs showed accuracies ranging from 92.0% to 100.0%. These results demonstrate that combining enterprise-domain knowledge retrieval with schema-constrained output generation can substantially improve the reliability of LLM-based generation for proprietary reservoir simulation software.
- Open Access
- 1 September 2026
- Published by KEAI Publishing in Petroleum Science
An integrated history-matching workflow for well monitoring and seismic data in geological CO2 sequestration applications
- 1 September 2026
- Published by KEAI Publishing in Petroleum Science
Open AccessAn integrated history-matching workflow for well monitoring and seismic data in geological CO2 sequestration applications
- 1 July 2026
- Published by AIP Publishing in Physics of Fluids
- Vol. 38  (7)
Quantifying CO2 huff-n-puff enhanced oil recovery in shale reservoirs via nuclear magnetic resonance experiments and compositional simulation
Shale oil reservoirs possess abundant resources yet face low primary recovery (<10%). CO2 huff-n-puff has attracted significant attention as a promising enhanced oil recovery (EOR) method for these reservoirs. However, based solely on experimental investigations, the underlying EOR mechanisms remain poorly understood. To address these challenges, we propose an integrated method combining experiment and numerical modeling to distinguish and quantify the mechanisms during CO2 huff-n-puff. Experimentally, we employed nuclear magnetic resonance to monitor real-time fluid migration within shale cores. Analysis of dynamic T2 spectra and one-dimensional frequency-encoded signals enabled calculation of real-time recovery rates and effective distances. Numerically, a core-scale compositional simulation model was built and calibrated to match experimental data. Experimental results show that CO2 effective distance is primarily governed by diffusion and limited to the centimeter scale, resulting in low oil recovery. The effective distance increases significantly with soaking time but is unaffected by injection pressure, indicating that pressure alone is insufficient to mobilize deeper oil. The calibrated model quantified the contributions of four mechanisms: oil swelling, diffusion, dissolved gas drive, and elastic energy release. Dissolved gas drive was identified as dominant, accounting for over 40% of total oil recovery. Parameter variations showed that higher injection pressure enhances diffusion of oil molecules into the CO2 phase, improving recovery via oil-molecule diffusion, while prolonged soaking time facilitates CO2 diffusion into the matrix, boosting recovery through oil swelling and dissolved gas drive. These findings provide mechanistic insights for optimizing CO2 huff-n-puff strategies and improving shale oil recovery efficiency.
- 1 July 2026
- Published by AIP Publishing in Physics of Fluids
- Vol. 38  (7)
Quantifying CO2 huff-n-puff enhanced oil recovery in shale reservoirs via nuclear magnetic resonance experiments and compositional simulation
Shale oil reservoirs possess abundant resources yet face low primary recovery (<10%). CO2 huff-n-puff has attracted significant attention as a promising enhanced oil recovery (EOR) method for these reservoirs. However, based solely on experimental investigations, the underlying EOR mechanisms remain poorly understood. To address these challenges, we propose an integrated method combining experiment and numerical modeling to distinguish and quantify the mechanisms during CO2 huff-n-puff. Experimentally, we employed nuclear magnetic resonance to monitor real-time fluid migration within shale cores. Analysis of dynamic T2 spectra and one-dimensional frequency-encoded signals enabled calculation of real-time recovery rates and effective distances. Numerically, a core-scale compositional simulation model was built and calibrated to match experimental data. Experimental results show that CO2 effective distance is primarily governed by diffusion and limited to the centimeter scale, resulting in low oil recovery. The effective distance increases significantly with soaking time but is unaffected by injection pressure, indicating that pressure alone is insufficient to mobilize deeper oil. The calibrated model quantified the contributions of four mechanisms: oil swelling, diffusion, dissolved gas drive, and elastic energy release. Dissolved gas drive was identified as dominant, accounting for over 40% of total oil recovery. Parameter variations showed that higher injection pressure enhances diffusion of oil molecules into the CO2 phase, improving recovery via oil-molecule diffusion, while prolonged soaking time facilitates CO2 diffusion into the matrix, boosting recovery through oil swelling and dissolved gas drive. These findings provide mechanistic insights for optimizing CO2 huff-n-puff strategies and improving shale oil recovery efficiency.
- 3 April 2026
- Published by Society of Petroleum Engineers (SPE) in SPE Journal
- Vol. 31  (06) , 4131-4146
Prediction of Sucker Rod Pump Failures Using Scaled Load Ratios and Machine Learning
Summary: Monitoring of sucker rod pumps is essential for maintaining continuous production. Traditionally, pump anomalies have been diagnosed by classifying downhole pump cards that are numerically calculated from surface measurements and rod string properties. However, pumps may continue to operate normally even when their downhole pump cards are classified as abnormal. Under constraints of limited manpower and budget, prioritizing the prediction and prevention of actual pump failures is more practical than attempting to manage all frequent anomalies detected through downhole pump card classification. With this study, we propose a real-time pump failure prediction method utilizing deep learning and a novel metric—the scaled load ratio, which is the ratio of the normalized minimum to maximum surface rod loads. The proposed method does not require downhole pump card computations, as it accurately predicts pump failures using only surface pump load data. In predicting pump failures in two United States (US) shale oil fields, the proposed method achieved an F1 score of 0.857 and was able to predict failures with an average lead time of 13.97 days before they occurred.
- 3 April 2026
- Published by Society of Petroleum Engineers (SPE) in SPE Journal
- Vol. 31  (06) , 4131-4146
Prediction of Sucker Rod Pump Failures Using Scaled Load Ratios and Machine Learning
Summary: Monitoring of sucker rod pumps is essential for maintaining continuous production. Traditionally, pump anomalies have been diagnosed by classifying downhole pump cards that are numerically calculated from surface measurements and rod string properties. However, pumps may continue to operate normally even when their downhole pump cards are classified as abnormal. Under constraints of limited manpower and budget, prioritizing the prediction and prevention of actual pump failures is more practical than attempting to manage all frequent anomalies detected through downhole pump card classification. With this study, we propose a real-time pump failure prediction method utilizing deep learning and a novel metric—the scaled load ratio, which is the ratio of the normalized minimum to maximum surface rod loads. The proposed method does not require downhole pump card computations, as it accurately predicts pump failures using only surface pump load data. In predicting pump failures in two United States (US) shale oil fields, the proposed method achieved an F1 score of 0.857 and was able to predict failures with an average lead time of 13.97 days before they occurred.
- 21 November 2025
- Published by Society of Petroleum Engineers (SPE) in SPE Journal
- Vol. 31  (01) , 75-94
Efficient Training of Deep-Learning-Based Surrogate Models for Subsurface Flow Simulation Using Multifidelity Data and Physics Constraints
Summary: In the field of subsurface flow simulation, deep-learning-based surrogate modeling is shown as a promising approach to significantly reduce the computational cost associated with full-physics reservoir simulations. However, the successful construction of highly accurate deep-learning-based models often requires a large number of training simulations, which can be very time-consuming to generate by itself for large-scale systems. The incorporation of physics constraints during the training process was shown to be an effective approach to reduce the training cost and to improve model accuracy. However, it remains challenging to reimplement the physics constraints correctly for modeling training when it comes to reservoir models with complex physics or gridding methods. In this study, we propose a physics-constrained surrogate (PCS) model training approach for production optimization that integrates multifidelity training data, together with a new way of implementing the physics constraints that leverages an existing mature reservoir simulator. The proposed approach starts with model pretraining on a relatively large number of low-cost, low-fidelity training data, followed by model fine-tuning with a smaller number of high-fidelity training samples and the inclusion of physics constraints. The physics constraints are implemented by adding the residuals of discretized governing equations into the loss function. Training with multifidelity data and incorporating physics constraints allows for reduced reliance on high-fidelity data while enhancing physical consistency. Systematic comparison studies were performed for both 2D and 3D cases, and it was shown that the proposed approach can reduce the computational costs associated with training data generation by about 80%, while achieving a similar level of prediction accuracy of the surrogate model. In addition, under a small number of training simulations (i.e., 50 equivalent high-fidelity runs), our proposed PCS model with multifidelity data can reduce the prediction error by 90%, in comparison with the model trained with only high-fidelity data. Finally, the trained surrogate model was applied to a well-control optimization problem. In comparison with the use of full-order simulations, the total computational time can be reduced by 97.7%.
- 21 November 2025
- Published by Society of Petroleum Engineers (SPE) in SPE Journal
- Vol. 31  (01) , 75-94
Efficient Training of Deep-Learning-Based Surrogate Models for Subsurface Flow Simulation Using Multifidelity Data and Physics Constraints
Summary: In the field of subsurface flow simulation, deep-learning-based surrogate modeling is shown as a promising approach to significantly reduce the computational cost associated with full-physics reservoir simulations. However, the successful construction of highly accurate deep-learning-based models often requires a large number of training simulations, which can be very time-consuming to generate by itself for large-scale systems. The incorporation of physics constraints during the training process was shown to be an effective approach to reduce the training cost and to improve model accuracy. However, it remains challenging to reimplement the physics constraints correctly for modeling training when it comes to reservoir models with complex physics or gridding methods. In this study, we propose a physics-constrained surrogate (PCS) model training approach for production optimization that integrates multifidelity training data, together with a new way of implementing the physics constraints that leverages an existing mature reservoir simulator. The proposed approach starts with model pretraining on a relatively large number of low-cost, low-fidelity training data, followed by model fine-tuning with a smaller number of high-fidelity training samples and the inclusion of physics constraints. The physics constraints are implemented by adding the residuals of discretized governing equations into the loss function. Training with multifidelity data and incorporating physics constraints allows for reduced reliance on high-fidelity data while enhancing physical consistency. Systematic comparison studies were performed for both 2D and 3D cases, and it was shown that the proposed approach can reduce the computational costs associated with training data generation by about 80%, while achieving a similar level of prediction accuracy of the surrogate model. In addition, under a small number of training simulations (i.e., 50 equivalent high-fidelity runs), our proposed PCS model with multifidelity data can reduce the prediction error by 90%, in comparison with the model trained with only high-fidelity data. Finally, the trained surrogate model was applied to a well-control optimization problem. In comparison with the use of full-order simulations, the total computational time can be reduced by 97.7%.
- 1 October 2025
- Published by Society of Petroleum Engineers (SPE) in SPE Journal
- Vol. 30  (12) , 7822-7839
Reduced-Order Modeling for Fractured Reservoir Simulation by Use of Local Resolution Trajectory Piecewise Linearization
Summary: Flow simulation for fractured reservoirs requires more computational cost than that in conventional reservoirs, especially when fractures need to be explicitly characterized. The objective of this study is to present a novel reduced-order model (ROM) approach for accelerating the flow simulation of fractured reservoirs. The methodology combines embedded discrete fracture modeling (EDFM) with a ROM approach based on proper orthogonal decomposition (POD) and trajectory piecewise linearization (TPWL). The POD-TPWL-based ROM can provide significant simulation speedup by linearizing the original governing equations and projecting the high-dimensional linearized representations into a low-dimensional space. However, the basic POD-TPWL-based ROM can be unstable and inaccurate when dealing with complex models with fractures. In this study, we developed a local resolution basis-matrix construction approach to better capture the multiscale flow dynamics in the matrix and fracture domain, especially for the grid cells perforated by wells. We tested the accuracy and robustness of the ROM with several numerical cases. The results showed that the proposed ROM maintains a high level of prediction accuracy in both matrix and fracture domains, even when the well control settings in test cases diverge substantially from those used in the training simulations. The speedup factor of the POD-TPWL-based ROM, compared with the full-order model (FOM), ranges from 100 to 1,000, depending on the model size and complexities. For a 3D fracture case, a speedup factor of 690 is achieved while maintaining relative error of cumulative production below 0.02 for all producers.
- 1 October 2025
- Published by Society of Petroleum Engineers (SPE) in SPE Journal
- Vol. 30  (12) , 7822-7839
Reduced-Order Modeling for Fractured Reservoir Simulation by Use of Local Resolution Trajectory Piecewise Linearization
Summary: Flow simulation for fractured reservoirs requires more computational cost than that in conventional reservoirs, especially when fractures need to be explicitly characterized. The objective of this study is to present a novel reduced-order model (ROM) approach for accelerating the flow simulation of fractured reservoirs. The methodology combines embedded discrete fracture modeling (EDFM) with a ROM approach based on proper orthogonal decomposition (POD) and trajectory piecewise linearization (TPWL). The POD-TPWL-based ROM can provide significant simulation speedup by linearizing the original governing equations and projecting the high-dimensional linearized representations into a low-dimensional space. However, the basic POD-TPWL-based ROM can be unstable and inaccurate when dealing with complex models with fractures. In this study, we developed a local resolution basis-matrix construction approach to better capture the multiscale flow dynamics in the matrix and fracture domain, especially for the grid cells perforated by wells. We tested the accuracy and robustness of the ROM with several numerical cases. The results showed that the proposed ROM maintains a high level of prediction accuracy in both matrix and fracture domains, even when the well control settings in test cases diverge substantially from those used in the training simulations. The speedup factor of the POD-TPWL-based ROM, compared with the full-order model (FOM), ranges from 100 to 1,000, depending on the model size and complexities. For a 3D fracture case, a speedup factor of 690 is achieved while maintaining relative error of cumulative production below 0.02 for all producers.
- Open Access
- 1 June 2025
- Published by KEAI Publishing in Petroleum Science
- Vol. 22  (6) , 2478-2492
A dual-model dual-grid upscaling method for solid-based thermal-reactive-compositional flow simulations in fractured oil shale reservoirs
- 1 June 2025
- Published by KEAI Publishing in Petroleum Science
- Vol. 22  (6) , 2478-2492
Open AccessA dual-model dual-grid upscaling method for solid-based thermal-reactive-compositional flow simulations in fractured oil shale reservoirs
- Top Cited PapersOpen Access
- 1 April 2025
- Published by KEAI Publishing in Petroleum Science
- Vol. 22  (4) , 1736-1756
Efficient deep-learning-based surrogate model for reservoir production optimization using transfer learning and multi-fidelity data
- 1 April 2025
- Published by KEAI Publishing in Petroleum Science
- Vol. 22  (4) , 1736-1756
Top Cited PapersOpen AccessEfficient deep-learning-based surrogate model for reservoir production optimization using transfer learning and multi-fidelity data
- 22 January 2025
- Published by Society of Petroleum Engineers (SPE) in SPE Journal
- Vol. 30  (03) , 1377-1395
A Deep Learning–Based Model for Real-Time Production Forecasting of Gas Wells with Intermittent Production
Summary: Accurate real-time production forecasting for gas wells experiencing intermittent production due to liquid loading is critical for optimizing operating schedules. Traditional methods, which rely on solving wellbore multiphase flow equations, are computationally intensive, making them impractical for real-time production forecasting across hundreds or thousands of wells, especially in the absence of reliable downhole measurements. This study presents an efficient real-time prediction model for production parameters of intermittent gas wells using deep learning techniques. The model generates production forecasts based on high-frequency measurements, including casing and tubing pressures, valve status, and instantaneous gas rates. To enhance the model’s performance, a novel cycle parameter extraction procedure is introduced during the feature engineering phase. This procedure extracts key performance indicators, specifically production parameters at critical moments, based on the shut-in and open cycles. Using these key cycle parameters as features significantly reduces the data volume used for model training and enhances prediction accuracy. The cycle parameter data set is derived from high-frequency measurements collected from 216 low-productivity gas wells. The developed deep learning model uses a Seq2Seq architecture. The encoding long short-term memory (LSTM) processes recent production data and passes the encoded information to the decoding LSTM, which then uses future operating schedules as inputs to predict the corresponding key cycle parameters. Testing results demonstrate the model’s high accuracy and efficiency, with each prediction taking less than 0.002 seconds. The prediction accuracy is quantified using the coefficient of determination (R²), the mean absolute error (MAE), and the mean absolute percentage error (MAPE). The R² values for the predicted features range from 91% to 99%, with a maximum MAE of 0.15 MPa for pressure parameters and 0.10×104 m³/d for the average gas rate (AGR). Moreover, the model maintains high accuracy even in challenging scenarios, such as wells with extended shut-in durations and insufficient measurements. The model’s predictive performance under various operating schedules is further evaluated using virtual cycles, demonstrating its strong generalization capability in generating production forecasts for unseen operating schedules in a given well. This generalization capability underscores the advantage of training the model with data collected from numerous wells operating under diverse schedules.
- 22 January 2025
- Published by Society of Petroleum Engineers (SPE) in SPE Journal
- Vol. 30  (03) , 1377-1395
A Deep Learning–Based Model for Real-Time Production Forecasting of Gas Wells with Intermittent Production
Summary: Accurate real-time production forecasting for gas wells experiencing intermittent production due to liquid loading is critical for optimizing operating schedules. Traditional methods, which rely on solving wellbore multiphase flow equations, are computationally intensive, making them impractical for real-time production forecasting across hundreds or thousands of wells, especially in the absence of reliable downhole measurements. This study presents an efficient real-time prediction model for production parameters of intermittent gas wells using deep learning techniques. The model generates production forecasts based on high-frequency measurements, including casing and tubing pressures, valve status, and instantaneous gas rates. To enhance the model’s performance, a novel cycle parameter extraction procedure is introduced during the feature engineering phase. This procedure extracts key performance indicators, specifically production parameters at critical moments, based on the shut-in and open cycles. Using these key cycle parameters as features significantly reduces the data volume used for model training and enhances prediction accuracy. The cycle parameter data set is derived from high-frequency measurements collected from 216 low-productivity gas wells. The developed deep learning model uses a Seq2Seq architecture. The encoding long short-term memory (LSTM) processes recent production data and passes the encoded information to the decoding LSTM, which then uses future operating schedules as inputs to predict the corresponding key cycle parameters. Testing results demonstrate the model’s high accuracy and efficiency, with each prediction taking less than 0.002 seconds. The prediction accuracy is quantified using the coefficient of determination (R²), the mean absolute error (MAE), and the mean absolute percentage error (MAPE). The R² values for the predicted features range from 91% to 99%, with a maximum MAE of 0.15 MPa for pressure parameters and 0.10×104 m³/d for the average gas rate (AGR). Moreover, the model maintains high accuracy even in challenging scenarios, such as wells with extended shut-in durations and insufficient measurements. The model’s predictive performance under various operating schedules is further evaluated using virtual cycles, demonstrating its strong generalization capability in generating production forecasts for unseen operating schedules in a given well. This generalization capability underscores the advantage of training the model with data collected from numerous wells operating under diverse schedules.
- Open Access
- 1 January 2025
- Published by KEAI Publishing in Petroleum Science
- Vol. 22  (1) , 193-209
Characterization of petrophysical and seismic properties for CO2 storage with sensitivity analysis
- 1 January 2025
- Published by KEAI Publishing in Petroleum Science
- Vol. 22  (1) , 193-209
Open AccessCharacterization of petrophysical and seismic properties for CO2 storage with sensitivity analysis