Furui Zhan
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
Furui Zhan has a total of 83 co-authors. Here are the top co-authors by number of shared publications:
- Qian Ma
Dalian Maritime University
10 shared publications - Chengwei Zhang
College of Information Science and Technology, Dalian Maritime University, Dalian, China
9 shared publications - Hui Li
Dalian Maritime University, Dalian, China
9 shared publications Shikai Guo
6 shared publications- Shikai Guo
Dalian Maritime University, Dalian, China
6 shared publications Wanting Liu
Dalian Maritime University, 1 Linghai Road, Dalian, 116026, Liaoning, China
6 shared publications
Latest Publications
- Open Access
- 18 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Software Engineering and Methodology
- Vol. 35 (12) , 1-40
JIT-DCK: A KAN Multi-Task Model for Just-In-Time Code Defect Prediction and Localization
In software quality assurance, Just-In-Time Defect Prediction (JIT-DP) and Just-In-Time Defect Localization (JIT-DL) automate the identification of potential defective code and localization of defect lines by analyzing program change characteristics and code commit contexts. These techniques enable just-in-time quality monitoring and defect prevention during development, while also reducing manual costs and improving reliability. Due to the high dynamism and unstructured noise in code change data within agile development scenarios, as well as the difficulty in capturing complex nonlinear synergistic effects of multidimensional expert features, existing work faces dual challenges: lack of data robustness and lack of fine-grained modeling of expert features. To address these challenges, we propose a multi-task interaction technology JIT-DCK , which contains three core components: defect feature preprocessing component, defect prediction component, and defect localization component. The defect feature preprocessing component creates a subword vocabulary, controls extreme values and removes interference to generate data with three aspects, sending the line-level structure to the defect localization component and the semantic aggregation and expert features to the defect prediction component. The defect prediction component is structured with a three-layer cooperative architecture. It works on fusing semantic and line-level features and also models the complex nonlinear relationships of 14-dimensional expert features to achieve effective defect prediction. The defect localization component builds a closed-loop decision chain. It integrates commit-level global features and line-level local semantics through multi-level encoding and dynamic attention mechanisms for accurately localizing defects. Experiments show that JIT-DCK significantly outperforms SOTA models in defect prediction and localization metrics such as F1-Score + 2.76%, AUC + 1.79%, CodeChurn@20%Recall + 2.88%, Top5-Acc + 3.10%, Top10-Acc + 2.55%, Recall@20%CodeChurn l + 6.46%, CodeChurn@20%Recall l + 18.58%, and IFA + 3.57%, demonstrating its competence in code defect prediction and localization tasks.
- 18 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Software Engineering and Methodology
- Vol. 35 (12) , 1-40
Open AccessJIT-DCK: A KAN Multi-Task Model for Just-In-Time Code Defect Prediction and Localization
In software quality assurance, Just-In-Time Defect Prediction (JIT-DP) and Just-In-Time Defect Localization (JIT-DL) automate the identification of potential defective code and localization of defect lines by analyzing program change characteristics and code commit contexts. These techniques enable just-in-time quality monitoring and defect prevention during development, while also reducing manual costs and improving reliability. Due to the high dynamism and unstructured noise in code change data within agile development scenarios, as well as the difficulty in capturing complex nonlinear synergistic effects of multidimensional expert features, existing work faces dual challenges: lack of data robustness and lack of fine-grained modeling of expert features. To address these challenges, we propose a multi-task interaction technology JIT-DCK , which contains three core components: defect feature preprocessing component, defect prediction component, and defect localization component. The defect feature preprocessing component creates a subword vocabulary, controls extreme values and removes interference to generate data with three aspects, sending the line-level structure to the defect localization component and the semantic aggregation and expert features to the defect prediction component. The defect prediction component is structured with a three-layer cooperative architecture. It works on fusing semantic and line-level features and also models the complex nonlinear relationships of 14-dimensional expert features to achieve effective defect prediction. The defect localization component builds a closed-loop decision chain. It integrates commit-level global features and line-level local semantics through multi-level encoding and dynamic attention mechanisms for accurately localizing defects. Experiments show that JIT-DCK significantly outperforms SOTA models in defect prediction and localization metrics such as F1-Score + 2.76%, AUC + 1.79%, CodeChurn@20%Recall + 2.88%, Top5-Acc + 3.10%, Top10-Acc + 2.55%, Recall@20%CodeChurn l + 6.46%, CodeChurn@20%Recall l + 18.58%, and IFA + 3.57%, demonstrating its competence in code defect prediction and localization tasks.
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 45 (11) , 5765-5778
Structural-Mutation-Based Differential Testing for FPGA Logic Synthesis Compilers
Field Programmable Gate Arrays (FPGAs) play a crucial role in Electronic Design Automation (EDA) applications, which have been widely used in safety-critical environments, including aerospace, chip manufacturing, and medical devices. A critical stage in FPGA development is logic synthesis, which enables developers to translate their software designs into hardware net lists. However, bugs in FPGA logic synthesis compilers may lead to unexpected behaviors in target hardware net lists, which may cause potential security risks. Therefore, it is critical to detect and repair such bugs in FPGA logic synthesis compilers. While recent testing methods have made progress in automatic bug detection, they remain limited by a fundamental weakness, the generated test cases exhibit minimal timing complexity and structural diversity. To address this challenge, we propose a mutation-based method integrated with Bayesian optimization to detect bugs in FPGA logic synthesis compilers. The method, called SmootHDL, consists with three key components, the Test-program Generation Component (TPG), the Bayesian Diversity Selection Component (BDS), and the Equivalent Test Check Component (ETC). SmootHDL starts with a set of test cases which are called original test cases. To expand the timing test space, the TPG use a control-flow mutation strategy to generate new test cases which is functional equivalent with the original test cases. Furthermore, for increasing diversity of the generated test cases, the BDS combine the Bayesian optimizer to construct a selector based on the control-flow diversity. Finally, the ETC identifies bugs by comparing simulation outputs between original and generated test cases. Since the mutations preserve functional equivalence, functional equivalent test case should have same output. When SmootHDL detects any discrepancy between the outputs, it indicates a potential compiler bug. SmootHDL then reproduces the inconsistent behavior and reports the bug to official developers. We evaluated our method on the mainstream FPGA logic synthesis compilers. Experiments show that SmootHDL can generate more diverse test cases and detect more bugs compared with the latest testing methods for FPGA logic synthesis compilers. Over three months, SmootHDL found 16 bugs, 12 of which were confirmed by official technical support, and 6 have already been fixed.
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 45 (11) , 5765-5778
Structural-Mutation-Based Differential Testing for FPGA Logic Synthesis Compilers
Field Programmable Gate Arrays (FPGAs) play a crucial role in Electronic Design Automation (EDA) applications, which have been widely used in safety-critical environments, including aerospace, chip manufacturing, and medical devices. A critical stage in FPGA development is logic synthesis, which enables developers to translate their software designs into hardware net lists. However, bugs in FPGA logic synthesis compilers may lead to unexpected behaviors in target hardware net lists, which may cause potential security risks. Therefore, it is critical to detect and repair such bugs in FPGA logic synthesis compilers. While recent testing methods have made progress in automatic bug detection, they remain limited by a fundamental weakness, the generated test cases exhibit minimal timing complexity and structural diversity. To address this challenge, we propose a mutation-based method integrated with Bayesian optimization to detect bugs in FPGA logic synthesis compilers. The method, called SmootHDL, consists with three key components, the Test-program Generation Component (TPG), the Bayesian Diversity Selection Component (BDS), and the Equivalent Test Check Component (ETC). SmootHDL starts with a set of test cases which are called original test cases. To expand the timing test space, the TPG use a control-flow mutation strategy to generate new test cases which is functional equivalent with the original test cases. Furthermore, for increasing diversity of the generated test cases, the BDS combine the Bayesian optimizer to construct a selector based on the control-flow diversity. Finally, the ETC identifies bugs by comparing simulation outputs between original and generated test cases. Since the mutations preserve functional equivalence, functional equivalent test case should have same output. When SmootHDL detects any discrepancy between the outputs, it indicates a potential compiler bug. SmootHDL then reproduces the inconsistent behavior and reports the bug to official developers. We evaluated our method on the mainstream FPGA logic synthesis compilers. Experiments show that SmootHDL can generate more diverse test cases and detect more bugs compared with the latest testing methods for FPGA logic synthesis compilers. Over three months, SmootHDL found 16 bugs, 12 of which were confirmed by official technical support, and 6 have already been fixed.
- Open Access
- 20 October 2026
- Published by Elsevier in SSRN Electronic Journal
Ensuring Individualized Inventory Safety in Multi-Agent Refined Oil Dispatching via Safety-Gated Modular Policy Optimization
Multi-agent reinforcement learning (MARL) offers a promising paradigm for coordinated refined oil dispatching, where each gas station acts as an agent making replenishment decisions under inventory safety constraints. A fundamental yet underexplored challenge is agent heterogeneity: gas stations exhibit widely varying demand volatility, yet existing constrained MARL methods impose a single global safety threshold with a unified Lagrange multiplier. We argue that this practice is structurally flawed: a uniform boundary either over-constrains high-volatility stations, forcing overly conservative policies, or under-constrains low-volatility ones, inviting stockout risks. To resolve this heterogeneity-safety conflict, we propose SAM-HAPPO. Its core idea is to replace the uniform threshold with individualized safety budgets derived from the safety-stock principle in inventory theory, scaling each agent's allowance by its demand coefficient of variation. To stabilize the dual-variable update, we further replace the standard integral-only Lagrangian update with a proportional-integral (PI) control law, which reduces constraint-induced multiplier oscillations while preserving the correct KKT fixed point. To translate these individualized constraints into decentralized execution, we equip the framework with a safety-gated modular architecture featuring dual routing branches, i.e., efficiency-driven and safety-driven branches, blended by a learnable risk gate, alongside a module specialization loss to prevent routing collapse. Experiments on real-world and synthetic dispatching scenarios demonstrate that SAM-HAPPO substantially reduces inventory violations while maintaining competitive dispatch efficiency and performs consistently under both weekday and holiday demand patterns.
- 20 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEnsuring Individualized Inventory Safety in Multi-Agent Refined Oil Dispatching via Safety-Gated Modular Policy Optimization
Multi-agent reinforcement learning (MARL) offers a promising paradigm for coordinated refined oil dispatching, where each gas station acts as an agent making replenishment decisions under inventory safety constraints. A fundamental yet underexplored challenge is agent heterogeneity: gas stations exhibit widely varying demand volatility, yet existing constrained MARL methods impose a single global safety threshold with a unified Lagrange multiplier. We argue that this practice is structurally flawed: a uniform boundary either over-constrains high-volatility stations, forcing overly conservative policies, or under-constrains low-volatility ones, inviting stockout risks. To resolve this heterogeneity-safety conflict, we propose SAM-HAPPO. Its core idea is to replace the uniform threshold with individualized safety budgets derived from the safety-stock principle in inventory theory, scaling each agent's allowance by its demand coefficient of variation. To stabilize the dual-variable update, we further replace the standard integral-only Lagrangian update with a proportional-integral (PI) control law, which reduces constraint-induced multiplier oscillations while preserving the correct KKT fixed point. To translate these individualized constraints into decentralized execution, we equip the framework with a safety-gated modular architecture featuring dual routing branches, i.e., efficiency-driven and safety-driven branches, blended by a learnable risk gate, alongside a module specialization loss to prevent routing collapse. Experiments on real-world and synthetic dispatching scenarios demonstrate that SAM-HAPPO substantially reduces inventory violations while maintaining competitive dispatch efficiency and performs consistently under both weekday and holiday demand patterns.
- Open Access
- 16 August 2026
- Published by Elsevier in SSRN Electronic Journal
Co-Design of Circuit Representations: Cross-Modal Graph Matching for Verilog-AIG Alignment
Circuit representations are essential for modeling large-scale digital circuits and facilitating efficient logic synthesis analysis. However, due to the inherent complexity of structural dependencies and functional interactions, existing approaches often struggle with inconsistencies between structural and functional behaviors. To address these issues, we propose AlignAIG, a novel framework for global circuit representation learning that ensures structural-functional consistency. Specifically, AlignAIG integrates a global message-passing network to capture long-range dependencies across large-scale circuits and employs a structured masking strategy that safeguards critical logic while selectively masking gates with minimal impact. Furthermore, AlignAIG introduces a cross-modal alignment decoder that utilizes attention mechanisms to align Verilog semantic features with circuit structural representations, enhancing the coherence between these two modalities. Extensive experiments on 65,500 circuit-structure files and 40,000 Verilog-AIG equivalence pairs across arithmetic, control and industrial domains demonstrate the effectiveness of AlignAIG. In the logic equivalence task, AlignAIG improves AUC by 17.01% and F1-score by 36.45% over state-of-the-art models on the global dataset. Additionally, it obtains an average gain of 39.7% in NDCG and 12.2% in the Top-𝑘% Commonality for QoR prediction. The source code is available at https: //github.com/zhongmoxingzhe/Cross_Modal_Graph_Matching_for_Verilog_AIG_Alignment.
- 16 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCo-Design of Circuit Representations: Cross-Modal Graph Matching for Verilog-AIG Alignment
Circuit representations are essential for modeling large-scale digital circuits and facilitating efficient logic synthesis analysis. However, due to the inherent complexity of structural dependencies and functional interactions, existing approaches often struggle with inconsistencies between structural and functional behaviors. To address these issues, we propose AlignAIG, a novel framework for global circuit representation learning that ensures structural-functional consistency. Specifically, AlignAIG integrates a global message-passing network to capture long-range dependencies across large-scale circuits and employs a structured masking strategy that safeguards critical logic while selectively masking gates with minimal impact. Furthermore, AlignAIG introduces a cross-modal alignment decoder that utilizes attention mechanisms to align Verilog semantic features with circuit structural representations, enhancing the coherence between these two modalities. Extensive experiments on 65,500 circuit-structure files and 40,000 Verilog-AIG equivalence pairs across arithmetic, control and industrial domains demonstrate the effectiveness of AlignAIG. In the logic equivalence task, AlignAIG improves AUC by 17.01% and F1-score by 36.45% over state-of-the-art models on the global dataset. Additionally, it obtains an average gain of 39.7% in NDCG and 12.2% in the Top-𝑘% Commonality for QoR prediction. The source code is available at https: //github.com/zhongmoxingzhe/Cross_Modal_Graph_Matching_for_Verilog_AIG_Alignment.
- Open Access
- 15 August 2026
- Published by Elsevier in SSRN Electronic Journal
ReconLight: Reconciling Individual Optimization and Global Coordination in Multi-Agent Traffic Signal Control
Multi-Agent Reinforcement Learning (MARL) has shown great promise for Adaptive Traffic Signal Control (ATSC). A core objective of ATSC is to minimize network-wide average travel time while satisfying practical constraints such as traffic fairness. However, achieving this in multi-intersection settings is difficult, primarily because the credit assignment problem makes it hard to attribute global outcomes to individual intersection decisions. To enable effective learning, existing methods resort to dense surrogate rewards, such as queue lengths or the number of waiting vehicles at individual intersections, which are immediately observable and readily optimized. However, there is no guaranty that these locally defined auxiliary objectives are consistently aligned with the global goal: reducing queue length at one intersection may simply shift congestion downstream rather than genuinely improve network-wide efficiency. Moreover, per-intersection surrogate rewards naturally overlook network-level concerns such as fairness, which requires balancing congestion across intersections and is inherently a global objective that cannot be captured by purely local metrics. To address these intertwined challenges, we propose ReconLight, a novel MARL framework that enables a smooth transition from locally responsive optimization to globally coordinated optimization through model aggregation across actor networks. ReconLight employs a dual actor-critic architecture that optimizes individual and group objectives separately, with a shared Graph Attention Network (GAT) capturing spatial dependencies among intersections. A time-varying parameter fusion mechanism progressively blends individual and global actor parameters using cosine-annealed weights and an initial-parameter anchor, enabling efficient and fair network-wide control through a stable model-aggregation process. Experiments on both synthetic and real-world datasets demonstrate that ReconLight consistently outperforms competitive baselines across all tested scenarios, achieving up to a 3.34\% reduction in average travel time.
- 15 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessReconLight: Reconciling Individual Optimization and Global Coordination in Multi-Agent Traffic Signal Control
Multi-Agent Reinforcement Learning (MARL) has shown great promise for Adaptive Traffic Signal Control (ATSC). A core objective of ATSC is to minimize network-wide average travel time while satisfying practical constraints such as traffic fairness. However, achieving this in multi-intersection settings is difficult, primarily because the credit assignment problem makes it hard to attribute global outcomes to individual intersection decisions. To enable effective learning, existing methods resort to dense surrogate rewards, such as queue lengths or the number of waiting vehicles at individual intersections, which are immediately observable and readily optimized. However, there is no guaranty that these locally defined auxiliary objectives are consistently aligned with the global goal: reducing queue length at one intersection may simply shift congestion downstream rather than genuinely improve network-wide efficiency. Moreover, per-intersection surrogate rewards naturally overlook network-level concerns such as fairness, which requires balancing congestion across intersections and is inherently a global objective that cannot be captured by purely local metrics. To address these intertwined challenges, we propose ReconLight, a novel MARL framework that enables a smooth transition from locally responsive optimization to globally coordinated optimization through model aggregation across actor networks. ReconLight employs a dual actor-critic architecture that optimizes individual and group objectives separately, with a shared Graph Attention Network (GAT) capturing spatial dependencies among intersections. A time-varying parameter fusion mechanism progressively blends individual and global actor parameters using cosine-annealed weights and an initial-parameter anchor, enabling efficient and fair network-wide control through a stable model-aggregation process. Experiments on both synthetic and real-world datasets demonstrate that ReconLight consistently outperforms competitive baselines across all tested scenarios, achieving up to a 3.34\% reduction in average travel time.
- Open Access
- 3 August 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
Architecture-Aware Multi-Scale Modeling with Self- and Cross-Attention for Dynamic IR Drop Analysis
With the continued advancement of Very Large-Scale Integration (VLSI) technology, IR drop in Power Delivery Networks (PDNs) poses a serious threat to chip performance and reliability. Due to the complex cross-scale interactions between local hotspots and global power paths in real-world scenarios, existing approaches face the challenges of lacking a coupled representation of multi-scale features and ignoring long-range dependencies and spatiotemporal feature fusion. This severely limits the ability of IR drop prediction approaches to capture voltage-drop hotspots, undermining the accuracy and reliability of their predictions. Therefore, this paper proposes a novel prediction framework named IR-Hunter. IR-Hunter first employs heterogeneous convolution kernels to model local electrothermal hotspots and global PDN impedance distributions, precisely capturing hotspot locations and macro-scale current-flow paths critical for chip layout and routing. Then, it adaptively fuses these multi-scale features. Additionally, IR-Hunter learns global power interactions within the PDN to identify long-range voltage coupling effects affecting routing and power-grid stability. Finally, IR-Hunter leverages cross-metal-layer current propagation pathways, efficiently integrating cross-scale spatiotemporal information between the encoder and decoder stages, consistently improving IR drop prediction accuracy during chip design. Comprehensive experiments conducted on 20,578 samples collected from over 20K instances demonstrate IR-Hunter’s superior prediction accuracy, achieving an average 4.57% NRMSE reduction and 4.12% SSIM improvement compared to state-of-the-art approaches. Additionally, to foster advancements in the EDA community, we have open-sourced the code at https://github.com/xhhlzy/IR-Hunter.
- 3 August 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
Open AccessArchitecture-Aware Multi-Scale Modeling with Self- and Cross-Attention for Dynamic IR Drop Analysis
With the continued advancement of Very Large-Scale Integration (VLSI) technology, IR drop in Power Delivery Networks (PDNs) poses a serious threat to chip performance and reliability. Due to the complex cross-scale interactions between local hotspots and global power paths in real-world scenarios, existing approaches face the challenges of lacking a coupled representation of multi-scale features and ignoring long-range dependencies and spatiotemporal feature fusion. This severely limits the ability of IR drop prediction approaches to capture voltage-drop hotspots, undermining the accuracy and reliability of their predictions. Therefore, this paper proposes a novel prediction framework named IR-Hunter. IR-Hunter first employs heterogeneous convolution kernels to model local electrothermal hotspots and global PDN impedance distributions, precisely capturing hotspot locations and macro-scale current-flow paths critical for chip layout and routing. Then, it adaptively fuses these multi-scale features. Additionally, IR-Hunter learns global power interactions within the PDN to identify long-range voltage coupling effects affecting routing and power-grid stability. Finally, IR-Hunter leverages cross-metal-layer current propagation pathways, efficiently integrating cross-scale spatiotemporal information between the encoder and decoder stages, consistently improving IR drop prediction accuracy during chip design. Comprehensive experiments conducted on 20,578 samples collected from over 20K instances demonstrate IR-Hunter’s superior prediction accuracy, achieving an average 4.57% NRMSE reduction and 4.12% SSIM improvement compared to state-of-the-art approaches. Additionally, to foster advancements in the EDA community, we have open-sourced the code at https://github.com/xhhlzy/IR-Hunter.
- Open Access
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
CircScale: Learning Scale-Adaptive Circuit Representations for Logic Synthesis Optimization
Logic Synthesis Optimization systematically transforms And-Inverter Graph (AIG) circuits through representation-guided transformations to improve circuit quality, serving as a critical step in Electronic Design Automation (EDA) workflows. However, real-world logic synthesis involves circuits with highly heterogeneous scales and dynamically evolving topologies. These characteristics pose significant challenges for previous circuit representation learning methods, which often struggle with the heterogeneous scale bottleneck and fail to capture the dynamic nature of optimization in static graph representations. This paper proposes CircScale, a framework that learns scale-adaptive circuit representations for logic synthesis optimization. Specifically, CircScale employs multi-hop neighborhood aggregation with circuit-aware gating to dynamically balance structural features across heterogeneous scales. It further introduces an early graph-sequence interaction mechanism that models circuit evolution and dynamically guides the optimization process. Comprehensive experiments are performed on 124,500 circuit-sequence pairs collected from five benchmarks, and the results demonstrate the superior representation capability of CircScale, achieving 14.89%-38.39% delay reduction and 23.89%-38.87% area improvement over state-of-the-art methods. These results confirm that CircScale effectively learns comprehensive and adaptive circuit representations for logic synthesis optimization. The source code is available at https://github.com/VNH0826/CircScale/.
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCircScale: Learning Scale-Adaptive Circuit Representations for Logic Synthesis Optimization
Logic Synthesis Optimization systematically transforms And-Inverter Graph (AIG) circuits through representation-guided transformations to improve circuit quality, serving as a critical step in Electronic Design Automation (EDA) workflows. However, real-world logic synthesis involves circuits with highly heterogeneous scales and dynamically evolving topologies. These characteristics pose significant challenges for previous circuit representation learning methods, which often struggle with the heterogeneous scale bottleneck and fail to capture the dynamic nature of optimization in static graph representations. This paper proposes CircScale, a framework that learns scale-adaptive circuit representations for logic synthesis optimization. Specifically, CircScale employs multi-hop neighborhood aggregation with circuit-aware gating to dynamically balance structural features across heterogeneous scales. It further introduces an early graph-sequence interaction mechanism that models circuit evolution and dynamically guides the optimization process. Comprehensive experiments are performed on 124,500 circuit-sequence pairs collected from five benchmarks, and the results demonstrate the superior representation capability of CircScale, achieving 14.89%-38.39% delay reduction and 23.89%-38.87% area improvement over state-of-the-art methods. These results confirm that CircScale effectively learns comprehensive and adaptive circuit representations for logic synthesis optimization. The source code is available at https://github.com/VNH0826/CircScale/.
- Open Access
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
DeepTempo: Learning Temporal Circuit Behaviors of Logic Gates
With the continuous evolution of integrated circuit technology toward higher complexity and larger scales, modern Electronic Design Automation (EDA) workflows demand sophisticated representation and analysis techniques to achieve optimal circuit design performance, power consumption, and reliability. And-Inverter Graphs (AIGs) serve as a fundamental structure for core EDA tasks, efficiently enabling logic synthesis, equivalence checking, and functional verification to enhance development efficiency. However, the growing complexity and integration density of modern circuits introduce increasingly intricate temporal signal propagation patterns and multi-layered feature interactions, which leads existing works ignoring temporal signal propagation dependencies and facing feature homogenization among deep aggregation in the accurate AIG representation learning. To address these challenges, we propose DeepTempo, a spatio-temporal joint modeling framework that effectively captures temporal signal propagation dependencies and preserves spatial features during deep aggregation to enable accurate AIG representation learning. Specifically, DeepTempo incorporates temporal convolution mechanisms to model sequential signal propagation, to capture the temporal dependencies inherent in circuit behavior. In addition, DeepTempo employs adaptive gating and cross-reconstruction strategies to preserve gate-specific functional distinctiveness during multi-layer feature aggregation, effectively mitigating feature homogenization problems to enhance discriminative modeling and achieve accurate AIG representation. We conduct comprehensive experiments on 9,933 valid circuit samples from multiple EDA benchmark domains covering logic synthesis, circuit testing, and hardware design to evaluate the effectiveness of DeepTempo in AIG representation learning. Experimental results on the Signal Probability Prediction (SPP) task and Truth-Table Distance Prediction (TTDP) task demonstrate that DeepTempo outperforms state-of-theart methods, achieving improvements of 14.29% (MAE) and 20.0% (MSE) on SPP, and achieving 21.35% (MAE) and 4.55% (MSE) on TTDP. We have open-sourced the project on GitHub to facilitate the advancement and development of defect prediction-related technologies .
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDeepTempo: Learning Temporal Circuit Behaviors of Logic Gates
With the continuous evolution of integrated circuit technology toward higher complexity and larger scales, modern Electronic Design Automation (EDA) workflows demand sophisticated representation and analysis techniques to achieve optimal circuit design performance, power consumption, and reliability. And-Inverter Graphs (AIGs) serve as a fundamental structure for core EDA tasks, efficiently enabling logic synthesis, equivalence checking, and functional verification to enhance development efficiency. However, the growing complexity and integration density of modern circuits introduce increasingly intricate temporal signal propagation patterns and multi-layered feature interactions, which leads existing works ignoring temporal signal propagation dependencies and facing feature homogenization among deep aggregation in the accurate AIG representation learning. To address these challenges, we propose DeepTempo, a spatio-temporal joint modeling framework that effectively captures temporal signal propagation dependencies and preserves spatial features during deep aggregation to enable accurate AIG representation learning. Specifically, DeepTempo incorporates temporal convolution mechanisms to model sequential signal propagation, to capture the temporal dependencies inherent in circuit behavior. In addition, DeepTempo employs adaptive gating and cross-reconstruction strategies to preserve gate-specific functional distinctiveness during multi-layer feature aggregation, effectively mitigating feature homogenization problems to enhance discriminative modeling and achieve accurate AIG representation. We conduct comprehensive experiments on 9,933 valid circuit samples from multiple EDA benchmark domains covering logic synthesis, circuit testing, and hardware design to evaluate the effectiveness of DeepTempo in AIG representation learning. Experimental results on the Signal Probability Prediction (SPP) task and Truth-Table Distance Prediction (TTDP) task demonstrate that DeepTempo outperforms state-of-theart methods, achieving improvements of 14.29% (MAE) and 20.0% (MSE) on SPP, and achieving 21.35% (MAE) and 4.55% (MSE) on TTDP. We have open-sourced the project on GitHub to facilitate the advancement and development of defect prediction-related technologies .
- Open Access
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Architecture-Aware Multi-Scale Modeling with Self-and Cross-Attention for Dynamic IR Drop Analysis
With the continued advancement of Very Large-Scale Integration (VLSI) technology, IR drop in Power Delivery Networks (PDNs) poses a serious threat to chip performance and reliability. Due to the complex cross-scale interactions between local hotspots and global power paths in real-world scenarios, existing approaches face the challenges of lacking a coupled representation of multi-scale features and ignoring long-range dependencies and spatiotemporal feature fusion. This severely limits the ability of IR drop prediction approaches to capture voltage-drop hotspots, undermining the accuracy and reliability of their predictions. Therefore, this paper proposes a novel prediction framework named IR-Hunter. IR-Hunter first employs heterogeneous convolution kernels to model local electrothermal hotspots and global PDN impedance distributions, precisely capturing hotspot locations and macro-scale current-flow paths critical for chip layout and routing. Then, it adaptively fuses these multi-scale features. Additionally, IR-Hunter learns global power interactions within the PDN to identify long-range voltage coupling effects affecting routing and power-grid stability. Finally, IR-Hunter leverages cross-metal-layer current propagation pathways, efficiently integrating cross-scale spatiotemporal information between encoder and decoder stages, consistently improving IR drop prediction accuracy during chip design. Comprehensive experiments conducted on 20,578 samples collected from over 20K instances demonstrate IR-Hunter’s superior prediction accuracy, achieving an average 4.57% NRMSE reduction and 4.12% SSIM improvement compared to state-of-the-art approaches. Additionally, to foster advancements in the EDA community, we have open-sourced the code at https://github.com/xhhlzy/IR-Hunter.
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessArchitecture-Aware Multi-Scale Modeling with Self-and Cross-Attention for Dynamic IR Drop Analysis
With the continued advancement of Very Large-Scale Integration (VLSI) technology, IR drop in Power Delivery Networks (PDNs) poses a serious threat to chip performance and reliability. Due to the complex cross-scale interactions between local hotspots and global power paths in real-world scenarios, existing approaches face the challenges of lacking a coupled representation of multi-scale features and ignoring long-range dependencies and spatiotemporal feature fusion. This severely limits the ability of IR drop prediction approaches to capture voltage-drop hotspots, undermining the accuracy and reliability of their predictions. Therefore, this paper proposes a novel prediction framework named IR-Hunter. IR-Hunter first employs heterogeneous convolution kernels to model local electrothermal hotspots and global PDN impedance distributions, precisely capturing hotspot locations and macro-scale current-flow paths critical for chip layout and routing. Then, it adaptively fuses these multi-scale features. Additionally, IR-Hunter learns global power interactions within the PDN to identify long-range voltage coupling effects affecting routing and power-grid stability. Finally, IR-Hunter leverages cross-metal-layer current propagation pathways, efficiently integrating cross-scale spatiotemporal information between encoder and decoder stages, consistently improving IR drop prediction accuracy during chip design. Comprehensive experiments conducted on 20,578 samples collected from over 20K instances demonstrate IR-Hunter’s superior prediction accuracy, achieving an average 4.57% NRMSE reduction and 4.12% SSIM improvement compared to state-of-the-art approaches. Additionally, to foster advancements in the EDA community, we have open-sourced the code at https://github.com/xhhlzy/IR-Hunter.
- Open Access
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
RL4HDL: Code Diversity Guided FPGA Logic Synthesis Compiler Testing Via Reinforcement Learning
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessRL4HDL: Code Diversity Guided FPGA Logic Synthesis Compiler Testing Via Reinforcement Learning