Research Logs
Research notes, experiments, and technical deep-dives on deep learning, optimization, applied AI, and engineering systems.
Penn State University / January 2026-March 2026
In this project, I focused on improving the detection of UNSAT cores within 3-SAT formulas, which are boolean equations where each clause contains three variables. My goal was to develop a method to identify these cores—the specific variables or contradictions that prevent an equation from being true—more quickly and efficiently than previous methods. To implement this, I generated training data containing UNSAT cores, visualized that data on a bipartite graph, and trained a Graphical Neural Network (GNN) to observe and improve how it interprets these clauses.During the research phase, I explored the complexities of the SAT algorithm and previous detection methods using resources like ScienceDirect and arXiv. Although I initially attempted to use a "Neural SAT" GitHub repository as a foundation, I found that much of the code was outdated and would not run locally. As a result, I had to build significant portions of the project from scratch, including developing my own scripts for data generation. Ultimately, my work combined visualization techniques and machine learning to better understand and solve contradictions within complex boolean logic.