Graph Neural Network for Boolean Satisfiability (SAT) solving using message passing on variable-clause graphs. Dataset for Article "Neural approaches to SAT solving: Design choices and interpretability" and PhD Thesis "Graph Neural Networks for Constraint Satisfaction: Theory, Design Space, and Connections to Continuous Relaxation Methods"

This record contains the dataset and associated code used in both the PhD thesis “Graph Neural Networks for Constraint Satisfaction: Theory, Design Space, and Connections to Continuous Relaxation Methods” by David Mojžíšek (ORCID: 0000-0002-3867-644X) and the journal article “Neural approaches to SAT solving: Design choices and interpretability” (DOI: 10.1016/j.ijar.2025.109609) by David Mojžíšek (ORCID: 0000-0002-3867-644X), Jan Hůla (ORCID: 0000-0001-7639-864X), Ziwei Li, Ziyu Zhou, and Mikoláš Janota (ORCID: 0000-0003-3487-784X).The dataset includes benchmark instances, generated problem sets, and outputs from graph neural network (GNN) models used for Boolean satisfiability (SAT) solving experiments. The code contains implementations for data generation, model training, evaluation scripts, and auxiliary utilities referenced in both the thesis and the article.In the published article, the authors provide a comprehensive evaluation of neural approaches to SAT solving, describe various design choices for GNN architectures, and analyze mechanisms for interpretability in reasoning over SAT instances. In the PhD thesis, these components are further contextualized within a broader theoretical and methodological framework of GNNs applied to constraint satisfaction problems, including connections to continuous relaxation methods.This record aims to ensure reproducibility and reuse of the dataset and code across the research community.

The journal article “Neural approaches to SAT solving: Design choices and interpretability” has been produced with the financial support of the European Union under the “Biography of Fake News with a Touch of AI: Dangerous Phenomenon through the Prism of Modern Human Sciences” project no. CZ.02.01.01/00/23_025/0008724 via the Operational Programme Jan Ámos Komenský. The research was also supported by the Ministry of Education, Youth and Sports within the dedicated program ERC CZ under the project POSTMAN no. LL1902, and by the Czech Science Foundation grant no. 25-17929X.