AutoPDR : Circuit-Aware Solver Configuration Prediction for Hardware Model Checking

Guangyu Hu 1 , Chen Chen 2 , Xiaofeng Zhou 1 , Jiaxi Zhang 3 , Wei Zhang 1 , Hongce Zhang 2 · 2026

Read AutoPDR : Circuit-Aware Solver Configuration Prediction for Hardware Model Checking by Guangyu Hu 1 , Chen Chen 2 , Xiaofeng Zhou 1 , Jiaxi Zhang 3 , Wei Zhang 1 , Hongce Zhang 2 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-25048.txt.

Introduction

Hu and co-authors automate solver setup for hardware model checking with AutoPDR. The PDR algorithm proves circuit safety, but its speed swings wildly with parameter choices. Hand tuning is slow and needs expert skill, while generic tuners waste time a proof task cannot spare. AutoPDR learns from circuit graphs plus traits of structure and wiring. It predicts a good setup for each circuit. Expert rules first drop bad options, cutting the search space by 78 percent. On a benchmark suite, the picked setups beat default and common settings.

Catalog record

CREATORGuangyu Hu 1 , Chen Chen 2 , Xiaofeng Zhou 1 , Jiaxi Zhang 3 , Wei Zhang 1 , Hongce Zhang 2
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34796
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.25048

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

AutoPDR : Circuit-Aware Solver Configuration Prediction for Hardware Model Checking Thanks: Guangyu Hu 1 , Chen Chen 2 , Xiaofeng Zhou 1 , Jiaxi Zhang 3 , Wei Zhang 1 , Hongce Zhang 2 Abstract Property Directed Reachability (PDR) is a powerful algorithm for formal verification of hardware and software systems, but its performance is highly sensitive to parameter configurations. Manual parameter tuning is time-consuming and requires domain expertise, while traditional automated parameter tuning frameworks are not well-suited for time-sensitive verification tasks like PDR. This paper presents a circuit-aware solver configuration framework that employs graph learning for intelligent heuristic selection in PDR-based verification. Our approach combines graph representations with static circuit features to predict optimal PDR solving configurations for specific circuits. We incorporate expert prior knowledge through constraint-based parameter filtering to eliminate invalid and inefficient configurations and reduce 78% search space. Our feature extraction pipeline captures structural, functional, and connectivity characteristics of circuit topology and component patterns. Experimental evaluation on a comprehensive benchmark suite demonstrates significant performance improvements compared to default configurations and commonly-used settings. The system successfully identifies circuit-specific parameter patterns and automatically selects the most suitable solving strategies based on circuit characteristics, making it a practical tool for automated formal verification workflows. The code is available at https://github.com/Gy-Hu/AutoPDR . I Introduction Formal verification has become an essential component of modern hardware and software design, providing mathematical guarantees about system correctness. Among the various verification techniques, Property Directed Reachability (PDR), also known as IC3 [ 2 ] , has emerged as one of the most effective algorithms for proving safety properties of finite-state systems. PDR’s success stems from its ability to incrementally construct inductive invariants while avoiding the explicit enumeration of reachable states, making it particularly well-suited for large-scale verification problems [ 9 , 6 ] . However, PDR’s effectiveness is highly dependent on its parameter configuration. One of the commonly-used PDR implementations is built in ABC [ 6 ] , which exposes numerous tunable parameters that control various aspects of its operation, including generalization strategies, abstraction mechanisms, proof obligation handling, and heuristic choices. These parameters significantly impact both the runtime performance and the likelihood of successful verification [ 7 , 4 ] . Unfortunately, finding optimal parameter settings is a challenging task that typically requires extensive domain expertise and time-consuming manual experimentation [ 10 ] . Moreover, the combinatorial explosion of parameter combinations makes parallel exploration computationally infeasible, as it would quickly exhaust available computational resources, such as CPU cores and memory. The parameter tuning problem in PDR is particularly complex due to several factors. First, the optimal parameter configuration is highly dependent on the structural characteristics of the circuit being verified. Parameters that work well for control-intensive designs may perform poorly on datapath-heavy circuits. Second, the parameter space is large and contains complex interdependencies—certain parameter combinations are logically invalid or lead to degenerate behavior. Third, the relationship between circuit features and optimal parameters is intricate and difficult to capture with traditional heuristics. These challenges highlight the need for a more systematic and learning-based approach that can capture the intricate relationships between circuit structure and parameter effectiveness. Current approaches to PDR parameter tuning fall into two categories: manual tuning by experts and simple automated search methods. Manual tuning is labor-intensive, as it requires deep understanding of both the algorithm and the target domain, and does not scale to diverse verification workloads. Automated search methods, such as random sampling or grid search, fail to exploit the structural information available in the verification problem and often waste computational resources exploring invalid or suboptimal configurations [ 11 , 14 ] . Furthermore, PDR engines frequently consume substantial memory resources, and solving times often exceed one hour, making such brute-force approaches impractical and unreliable. Recent advances in solver configuration [ 17 ] and circuit representation learning [ 16 , 19 ] suggest that machine learning offers a promising direction for addressing these challenges effectively. This paper presents a circuit-aware solver configuration framework that employs machine learning for intelligent heuristic selection in PDR-based verification. Our approach addresses the limitations of existing methods through three key innovations: • Circuit-Aware Feature Extraction: We develop a comprehensive feature extraction pipeline that captures both graph-level structural properties of And-Inverter Graph (AIG) representations and circuit-level statistical features. This multi-level characterization enables our models to understand the verification complexity and guide parameter selection accordingly. • Expert-Knowledge-Based Parameter Filtering: We incorporate human expertise through a systematic constraint-based approach to eliminate invalid parameter combinations based on logical dependencies among PDR options. This filtering mechanism significantly reduces the search space while ensuring that all considered configurations are semantically meaningful. • GNN-Based PDR Parameter Prediction: We conduct a comprehensive empirical evaluation of multiple graph neural network architectures, including GraphSAGE [ 8 ] , GIN [ 18 ] , GCN [ 12 ] , HOGA [ 5 ] and GraphSAINT [ 20 ] , to learn the complex mapping between circuit structure and optimal parameters. Through systematic comparison, we identify the most effective architecture for capturing both local node patterns and global graph properties that influence verification performance. Our approach builds upon recent advances in PDR algorithms [ 15 , 13 , 1 ] to optimize parameter selection. Our experimental evaluation on a comprehensive benchmark suite demonstrates significant improvements over existing approaches. Our framework increases the number of solved instances by 90.0% over the default PDR configuration and by 33.3% over a commonly-used optimized setting. The system successfully identifies circuit-specific optimization patterns and generalizes well across different verification domains. The main contributions of this work are: • A comprehensive machine learning framework for automated PDR parameter tuning that combines circuit feature extraction, parameter filtering, and graph neural network prediction. • A systematic parameter filtering mechanism that eliminates invalid configurations based on logical dependencies, significantly reducing the search space. • A thorough empirical evaluation of multiple GNN architectures, identifying the most effective approach for circuit-aware solver configuration. The remainder of this paper is organized as follows: Section II provides background on PDR algorithms and parameter tuning challenges. Section III presents our circuit-aware framework including feature extraction, parameter filtering, and GNN-based prediction. Section IV describes our experimental setup and evaluation methodology, including presenting and analyzing the experimental results. Section V concludes the paper and discusses future work. II Preliminaries II-A Property Directed Reachability (PDR/IC3) Property Directed Reachability (PDR) [ 2 ] , also known as IC3, is a SAT-based model checking algorithm that has revolutionized the formal verification of safety properties. Unlike traditional approaches such as Bounded Model Checking (BMC) or fixed-point iteration over symbolic states, PDR constructs a sequence of inductive, relative over-approximations of reachable states in a manner directed by the property being verified. The algorithm maintains a sequence of formulas F_{0},F_{1},\ldots,F_{k} , where each F_{i} represents an over-approximation of states reachable in at most i steps from the initial states. PDR incrementally strengthens these formulas by finding and generalizing counterexamples to induction (CTIs). This incremental and goal-oriented approach has made PDR exceptionally effective, particularly in hardware verification [ 6 ] . The performance of the PDR engine in the ABC verification tool, however, is critically dependent on a set of heuristic parameters that control its core operations. Understanding these parameters requires deep knowledge

Read the complete text