An AI-Assisted Logic Synthesis Framework: GNN-Based Quality-of-Results Prediction with Synthesis Recipe Recommendation for Accessible Chip Design

This work presents an AI-assisted logic synthesis framework for predicting synthesis Quality of Results (QoR) and supporting synthesis recipe selection for accessible chip design. The framework analyzes circuit structural characteristics and evaluates machine-learning-based QoR prediction using Linear Regression, Random Forest, and Gradient Boosting baselines. Experiments were conducted on a dataset containing 250 synthesis records generated from multiple circuits and synthesis recipes, with ten structural features used as model inputs. The models were evaluated on a held-out test set of 38 samples using MAE, MSE, RMSE, R², and MAPE. Among the evaluated baseline models, Linear Regression achieved the lowest RMSE of 0.179351. The framework also analyzes QoR variation across synthesis recipes and supports identification of high-performing synthesis configurations. The implementation includes data processing, model training, evaluation, automated result generation, feature analysis, and reproducibility tests, with 173 automated tests passing successfully.