Yukun Dong
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Yukun Dong has a total of 85 co-authors. Here are the top co-authors by number of shared publications:
- Shanchen Pang
China University of Petroleum (East China)
5 shared publications Shuai Liu
Qingdao Institute of Software, College of Computer Science and Technology, China University of Petroleum (East China), Qingdao, Shandong 266580, P. R. China
4 shared publicationsShuqi Wang
4 shared publicationsXiaoshan Liu
Qingdao Institute of Software, College of Computer Science and Technology, China University of Petroleum (East China), Qingdao, Shandong 266580, P. R. China
4 shared publicationsXiaotong Cheng
Qingdao Institute of Software, College of Computer Science and Technology, China University of Petroleum (East China), 66 Changjiang West Road, Huangdao District, Qingdao, Shandong Province, China
4 shared publications- Haojie Li
College of Computer Science and Technology, China University of Petroleum, Qingdao 266580, China
3 shared publications
Latest Publications
- Open Access
A Static Data-Driven Approach for Predicting the Dynamic Distribution of Physical Fields in Coalbed Methane Reservoirs with Multi-Fractured Horizontal Well
Accurately predicting the dynamic distribution of physical fields (including the gas saturation field, fracture pressure field, and matrix pressure field) in coalbed methane (CBM) reservoirs not only enables monitoring reservoir changes, but also provides support for refined reservoir management, which is of great significance for achieving efficient development and sustainable utilization of CBM reservoirs. However, existing research mainly focuses on CBM production prediction or pore pressure change analysis, and there are few studies on the prediction of the dynamic distribution of physical fields in CBM reservoirs. The systematic prediction of the dynamic distribution of physical fields is still in the exploratory stage. To address this research gap, this paper introduces a deep learning approach that utilizes static data to predict the dynamic distribution of physical fields. This method takes reservoir porosity, permeability, coal seam thickness, adsorption and desorption parameters, reservoir pressure, and other static data (collectively referred to as reservoir geological parameters), as well as multi-stage fracture information as input, and builds the pretraining model for multi-source feature fusion (MSFF-PM) and the multi-scale integrated convolutional long short-term memory (ConvLSTM) model (MSI-ConvLSTM). During inference, the framework predicts the dynamic distribution of physical fields from static reservoir parameters and fracture information without requiring historical field maps as external inputs. The coefficients of determination (R2) for gas saturation, fracture pressure, and matrix pressure predictions over the first five years were 0.953, 0.944, and 0.946, respectively. These findings are limited to the evaluated synthetic cases with homogenized background properties and idealized planar bi-wing fractures. These results demonstrate agreement with the CMG GEM reference fields within the evaluated synthetic cases.
Open AccessA Static Data-Driven Approach for Predicting the Dynamic Distribution of Physical Fields in Coalbed Methane Reservoirs with Multi-Fractured Horizontal Well
Accurately predicting the dynamic distribution of physical fields (including the gas saturation field, fracture pressure field, and matrix pressure field) in coalbed methane (CBM) reservoirs not only enables monitoring reservoir changes, but also provides support for refined reservoir management, which is of great significance for achieving efficient development and sustainable utilization of CBM reservoirs. However, existing research mainly focuses on CBM production prediction or pore pressure change analysis, and there are few studies on the prediction of the dynamic distribution of physical fields in CBM reservoirs. The systematic prediction of the dynamic distribution of physical fields is still in the exploratory stage. To address this research gap, this paper introduces a deep learning approach that utilizes static data to predict the dynamic distribution of physical fields. This method takes reservoir porosity, permeability, coal seam thickness, adsorption and desorption parameters, reservoir pressure, and other static data (collectively referred to as reservoir geological parameters), as well as multi-stage fracture information as input, and builds the pretraining model for multi-source feature fusion (MSFF-PM) and the multi-scale integrated convolutional long short-term memory (ConvLSTM) model (MSI-ConvLSTM). During inference, the framework predicts the dynamic distribution of physical fields from static reservoir parameters and fracture information without requiring historical field maps as external inputs. The coefficients of determination (R2) for gas saturation, fracture pressure, and matrix pressure predictions over the first five years were 0.953, 0.944, and 0.946, respectively. These findings are limited to the evaluated synthetic cases with homogenized background properties and idealized planar bi-wing fractures. These results demonstrate agreement with the CMG GEM reference fields within the evaluated synthetic cases.
- Open Access
- 8 November 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (21) , 11049
Cluster-Level Evidence-Enhanced Truth Inference for Knowledge Graph Fusion
Factual conflicts after entity alignment in knowledge graph fusion (KGF) can affect the consistency and reliability of fusion results. To address the difficulty of directly transferring existing traditional truth inference and fact verification methods to knowledge graph fusion, a cluster-level evidence-enhanced truth-inference method for knowledge graph fusion (CETI-KGF) is proposed. The method adopts a two-stage fact modeling strategy to organize semantically related multi-source facts and perform truth inference at the cluster level. Semantic-aware fact clustering (SFC) is constructed to divide facts into different fact groups. A cluster-level credibility evidence modeling (CEM) module is then employed to construct cluster-level evidence based on source support, count-based support, and local consistency within the knowledge graph. Finally, structured prompts guide the large language model (LLM) to jointly infer truths in a mixed single-truth and multi-truth setting, and ultimately map the resulting decisions to the facts within each cluster. The FR-EN subset of the entity alignment (EA) dataset DBP15K is used as the experimental setting, and a dataset containing 5429 conflict groups and 17,067 candidate facts is constructed for experimentation. The proposed method achieves the highest Error Precision and Error F1 among the compared methods, reaching 81.0% and 75.7%, respectively. Compared with directly using an LLM, the proposed method improves Error Precision and Error F1 by 10.5 and 6.0 percentage points, respectively.
- 8 November 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (21) , 11049
Open AccessCluster-Level Evidence-Enhanced Truth Inference for Knowledge Graph Fusion
Factual conflicts after entity alignment in knowledge graph fusion (KGF) can affect the consistency and reliability of fusion results. To address the difficulty of directly transferring existing traditional truth inference and fact verification methods to knowledge graph fusion, a cluster-level evidence-enhanced truth-inference method for knowledge graph fusion (CETI-KGF) is proposed. The method adopts a two-stage fact modeling strategy to organize semantically related multi-source facts and perform truth inference at the cluster level. Semantic-aware fact clustering (SFC) is constructed to divide facts into different fact groups. A cluster-level credibility evidence modeling (CEM) module is then employed to construct cluster-level evidence based on source support, count-based support, and local consistency within the knowledge graph. Finally, structured prompts guide the large language model (LLM) to jointly infer truths in a mixed single-truth and multi-truth setting, and ultimately map the resulting decisions to the facts within each cluster. The FR-EN subset of the entity alignment (EA) dataset DBP15K is used as the experimental setting, and a dataset containing 5429 conflict groups and 17,067 candidate facts is constructed for experimentation. The proposed method achieves the highest Error Precision and Error F1 among the compared methods, reaching 81.0% and 75.7%, respectively. Compared with directly using an LLM, the proposed method improves Error Precision and Error F1 by 10.5 and 6.0 percentage points, respectively.
- 17 October 2026
- Published by World Scientific Pub Co Pte Ltd in International Journal of Software Engineering and Knowledge Engineering
- p. 1-34
A Path-Guided and Iterative Optimization Strategy for Unit Test Case Generation with Large Language Models
The rise of Large Language Models (LLMs) has significantly impacted software engineering, particularly in automating unit test case generation. However, current methods often fall short of achieving comprehensive test coverage, especially for complex code structures and critical execution paths. This paper presents a novel Path-Guided and Iterative Optimization strategy for unit test case generation. By integrating LLMs with Control Flow Graph (CFG) analysis and iterative refinement, the strategy efficiently identifies untested execution paths, improves test coverage, and enhances test case accuracy. The proposed strategy employs a Path-Guided method, utilizing CFG analysis to explore reachable basic paths. It integrates Path-Guided contextual information with LLMs and applies Context-Prompt-Response (C–P–R) templated guidance to generate high-quality unit test cases. A Dual Validation mechanism ensures the assertion accuracy and reliability of the initial test cases, followed by Iterative Optimization, which progressively refines test cases to address coverage gaps. The experimental results demonstrate that the proposed strategy significantly enhances the quality of generated test cases. Using the Path-Guided and Iterative Optimization strategy, the method achieves a maximum of 97% line coverage and 94% branch coverage while also improving the precision of test cases.
- 17 October 2026
- Published by World Scientific Pub Co Pte Ltd in International Journal of Software Engineering and Knowledge Engineering
- p. 1-34
A Path-Guided and Iterative Optimization Strategy for Unit Test Case Generation with Large Language Models
The rise of Large Language Models (LLMs) has significantly impacted software engineering, particularly in automating unit test case generation. However, current methods often fall short of achieving comprehensive test coverage, especially for complex code structures and critical execution paths. This paper presents a novel Path-Guided and Iterative Optimization strategy for unit test case generation. By integrating LLMs with Control Flow Graph (CFG) analysis and iterative refinement, the strategy efficiently identifies untested execution paths, improves test coverage, and enhances test case accuracy. The proposed strategy employs a Path-Guided method, utilizing CFG analysis to explore reachable basic paths. It integrates Path-Guided contextual information with LLMs and applies Context-Prompt-Response (C–P–R) templated guidance to generate high-quality unit test cases. A Dual Validation mechanism ensures the assertion accuracy and reliability of the initial test cases, followed by Iterative Optimization, which progressively refines test cases to address coverage gaps. The experimental results demonstrate that the proposed strategy significantly enhances the quality of generated test cases. Using the Path-Guided and Iterative Optimization strategy, the method achieves a maximum of 97% line coverage and 94% branch coverage while also improving the precision of test cases.
- Open Access
- 8 October 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (19) , 9934
Library-Oriented Code Generation Method Combining Execution-Feedback Re-Retrieval and Repair Mechanisms
Library-oriented code generation methods typically supplement the library knowledge of large language models by retrieving API documentation. However, static retrieval cannot readily exploit runtime information generated during code generation to update the retrieved documents, and models may therefore still produce invalid API calls or code with runtime errors. To address this problem, we propose EFRRCoder, a library-oriented code generation method that combines execution-feedback-based re-retrieval and repair. On top of initial document retrieval and code generation, EFRRCoder introduces self-generated execution inputs to obtain runtime feedback from candidate programs, dynamically updates relevant API documents when execution reveals a potential API knowledge gap, and iteratively repairs failed code. This allows runtime information to improve generation results without additional model training. Experiments on the MonkeyEval and BeatNumEval library-oriented code generation benchmarks show that, across two base models, EFRRCoder improves pass@1 and pass@5 by up to 7.13 and 6.93 percentage points, respectively, over the best-performing baseline under each setting. Supplementary cross-library experiments on NdonnxEval further show that EFRRCoder consistently outperforms the compared baselines across both base models under docstring-removed and docstring-retained settings.
- 8 October 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (19) , 9934
Open AccessLibrary-Oriented Code Generation Method Combining Execution-Feedback Re-Retrieval and Repair Mechanisms
Library-oriented code generation methods typically supplement the library knowledge of large language models by retrieving API documentation. However, static retrieval cannot readily exploit runtime information generated during code generation to update the retrieved documents, and models may therefore still produce invalid API calls or code with runtime errors. To address this problem, we propose EFRRCoder, a library-oriented code generation method that combines execution-feedback-based re-retrieval and repair. On top of initial document retrieval and code generation, EFRRCoder introduces self-generated execution inputs to obtain runtime feedback from candidate programs, dynamically updates relevant API documents when execution reveals a potential API knowledge gap, and iteratively repairs failed code. This allows runtime information to improve generation results without additional model training. Experiments on the MonkeyEval and BeatNumEval library-oriented code generation benchmarks show that, across two base models, EFRRCoder improves pass@1 and pass@5 by up to 7.13 and 6.93 percentage points, respectively, over the best-performing baseline under each setting. Supplementary cross-library experiments on NdonnxEval further show that EFRRCoder consistently outperforms the compared baselines across both base models under docstring-removed and docstring-retained settings.
- 27 August 2026
- Published by World Scientific Pub Co Pte Ltd in International Journal of Software Engineering and Knowledge Engineering
- p. 1-33
VulRESC: A Vulnerability Detection Framework Based on Risk Path Extraction and Inter-Procedural Semantic Completion
Software vulnerability detection increasingly relies on learning-based models. However, most existing methods analyze individual functions in isolation, making it difficult to capture vulnerabilities caused by cross-function calls; directly introducing complete call chains can also lead to context expansion and noise accumulation. This paper proposes VulRESC, a vulnerability detection framework based on risk path extraction and inter-procedural semantic completion. The method first constructs code property graphs and extracts call paths related to external inputs and high-risk operations through a risk-point-driven pruning strategy. It then uses large language models to generate structured summaries for callees along the sequence in a bottom-up manner and introduces a variable-name alignment mechanism to bridge the semantic gap across function boundaries. Finally, the proposed DualVulBERT dual-stream model is employed to jointly identify vulnerabilities by fusing source code features and inter-procedural semantic summaries through cross-modal attention. Experimental results on real-world C/C[Formula: see text] vulnerability datasets show that VulRESC achieves an accuracy of 68.03% and an F1-score of 69.37%, outperforming representative existing methods.
- 27 August 2026
- Published by World Scientific Pub Co Pte Ltd in International Journal of Software Engineering and Knowledge Engineering
- p. 1-33
VulRESC: A Vulnerability Detection Framework Based on Risk Path Extraction and Inter-Procedural Semantic Completion
Software vulnerability detection increasingly relies on learning-based models. However, most existing methods analyze individual functions in isolation, making it difficult to capture vulnerabilities caused by cross-function calls; directly introducing complete call chains can also lead to context expansion and noise accumulation. This paper proposes VulRESC, a vulnerability detection framework based on risk path extraction and inter-procedural semantic completion. The method first constructs code property graphs and extracts call paths related to external inputs and high-risk operations through a risk-point-driven pruning strategy. It then uses large language models to generate structured summaries for callees along the sequence in a bottom-up manner and introduces a variable-name alignment mechanism to bridge the semantic gap across function boundaries. Finally, the proposed DualVulBERT dual-stream model is employed to jointly identify vulnerabilities by fusing source code features and inter-procedural semantic summaries through cross-modal attention. Experimental results on real-world C/C[Formula: see text] vulnerability datasets show that VulRESC achieves an accuracy of 68.03% and an F1-score of 69.37%, outperforming representative existing methods.
- 27 November 2025
- Published by Elsevier in Information and Software Technology
- Vol. 191
VDMPAGR: A vulnerability detection model based on pointer analysis and graph representation
- 27 November 2025
- Published by Elsevier in Information and Software Technology
- Vol. 191
VDMPAGR: A vulnerability detection model based on pointer analysis and graph representation
- Open Access
- 1 April 2025
- Published by KEAI Publishing in Natural Gas Industry B
- Vol. 12 (2) , 219-233
A novel surrogate model with deep learning for predicting spacial-temporal pressure in coalbed methane reservoirs
- 1 April 2025
- Published by KEAI Publishing in Natural Gas Industry B
- Vol. 12 (2) , 219-233
Open AccessA novel surrogate model with deep learning for predicting spacial-temporal pressure in coalbed methane reservoirs
- Open Access
Three-Dimensional Human Body Reconstruction Using Dual-View Normal Maps
High-fidelity 3D human body reconstruction is challenging, as single-view methods often lead to distortions due to self-occlusion, and the existing multi-view approaches either focus on pose or exhibit limited performance. This study presents an efficient approach to realistic 3D human body reconstruction from front and back images, emphasizing symmetry and surface detail preservation. We begin by extracting the key points and pose information from dual-view images, applying SMPL-X to generate an initial 3D body. Then, using normal maps derived from both views, we infer high-fidelity surfaces and optimize SMPL-X based on these reconstructed surfaces. Through implicit modeling, we merge the front and back surfaces, ensuring a symmetric fusion boundary for a complete 3D body model. Our experimental results on the THuman2.0 dataset demonstrate our method’s effectiveness, with significant improvements in the surface detail fidelity. To validate the model’s accuracy further, we collected waist and chest circumference measurements from 120 individuals, finding an average measurement error below 0.8 centimeters, thus confirming the robustness of SMPL-X optimized with dual-view data.
Open AccessThree-Dimensional Human Body Reconstruction Using Dual-View Normal Maps
High-fidelity 3D human body reconstruction is challenging, as single-view methods often lead to distortions due to self-occlusion, and the existing multi-view approaches either focus on pose or exhibit limited performance. This study presents an efficient approach to realistic 3D human body reconstruction from front and back images, emphasizing symmetry and surface detail preservation. We begin by extracting the key points and pose information from dual-view images, applying SMPL-X to generate an initial 3D body. Then, using normal maps derived from both views, we infer high-fidelity surfaces and optimize SMPL-X based on these reconstructed surfaces. Through implicit modeling, we merge the front and back surfaces, ensuring a symmetric fusion boundary for a complete 3D body model. Our experimental results on the THuman2.0 dataset demonstrate our method’s effectiveness, with significant improvements in the surface detail fidelity. To validate the model’s accuracy further, we collected waist and chest circumference measurements from 120 individuals, finding an average measurement error below 0.8 centimeters, thus confirming the robustness of SMPL-X optimized with dual-view data.
- 27 September 2024
- Published by Society of Petroleum Engineers (SPE) in SPE Journal
- Vol. 29 (11) , 5761-5778
Multisource Modeling Method for Petroleum Production Prediction Models: A Case Study of CO2-Flooding Performance
Summary: Accurate prediction of oil production is crucial for formulating oilfield development strategies. With the rapid development of artificial intelligence, research on utilizing deep learning to construct oil production prediction models has been growing, which has partially compensated for the low computational efficiency of numerical simulators. Although the well-trained source domain model maintains high prediction accuracy on target blocks with similar production conditions, the prediction accuracy of the model declines in scenarios where substantial disparities exist between the production conditions of the target block and the source domain. This discrepancy makes the prediction results unreliable and causes a domain shift issue. We propose a multisource model fine-tuning approach, which leverages a limited amount of target domain data to fine-tune the existing source domain model, enabling it to rapidly converge in the target domain while maintaining superior prediction performance. Based on a heterogeneous low-permeability CO2-flooding reservoir development model, we established a series of source domain data sets, encompassing numerous types of well patterns and permeability fields, and specifically prepared various target domain data sets to verify the effectiveness of the model fine-tuning. Experimental outcomes demonstrate that our proposed model fine-tuning approach facilitates the rapid convergence of the existing model on target domain data. Following testing, the fine-tuned model, which attained a prediction accuracy exceeding 97% in the target domain, significantly improved upon the accuracy compared with the unfine-tuned model. The time required is significantly lower than retraining a new model, and it significantly reduces the need for data in the target domain. This provides support for the rapid generation of new prediction models using existing source domain models and limited target domain data.
- 27 September 2024
- Published by Society of Petroleum Engineers (SPE) in SPE Journal
- Vol. 29 (11) , 5761-5778
Multisource Modeling Method for Petroleum Production Prediction Models: A Case Study of CO2-Flooding Performance
Summary: Accurate prediction of oil production is crucial for formulating oilfield development strategies. With the rapid development of artificial intelligence, research on utilizing deep learning to construct oil production prediction models has been growing, which has partially compensated for the low computational efficiency of numerical simulators. Although the well-trained source domain model maintains high prediction accuracy on target blocks with similar production conditions, the prediction accuracy of the model declines in scenarios where substantial disparities exist between the production conditions of the target block and the source domain. This discrepancy makes the prediction results unreliable and causes a domain shift issue. We propose a multisource model fine-tuning approach, which leverages a limited amount of target domain data to fine-tune the existing source domain model, enabling it to rapidly converge in the target domain while maintaining superior prediction performance. Based on a heterogeneous low-permeability CO2-flooding reservoir development model, we established a series of source domain data sets, encompassing numerous types of well patterns and permeability fields, and specifically prepared various target domain data sets to verify the effectiveness of the model fine-tuning. Experimental outcomes demonstrate that our proposed model fine-tuning approach facilitates the rapid convergence of the existing model on target domain data. Following testing, the fine-tuned model, which attained a prediction accuracy exceeding 97% in the target domain, significantly improved upon the accuracy compared with the unfine-tuned model. The time required is significantly lower than retraining a new model, and it significantly reduces the need for data in the target domain. This provides support for the rapid generation of new prediction models using existing source domain models and limited target domain data.
- 20 September 2024
- Published by Elsevier in Information and Software Technology
- Vol. 177
A search-and-fill strategy to code generation for complex software requirements
- 20 September 2024
- Published by Elsevier in Information and Software Technology
- Vol. 177
A search-and-fill strategy to code generation for complex software requirements