A Phase-Aware Automated Parameter Search Benchmark Framework and Search Behavior Study for OpenROAD Physical Design

Parameter settings in open-source EDA flows increasingly affect design convergence and physical results, making a reproducible and comparable automated parameter search framework necessary. Based on OpenROAD and OpenROAD-flow-scripts, this paper builds an automated parameter search benchmark that executes flows, parses logs, extracts metrics, and classifies trials into three categories: final, grt_partial, and invalid. Controlled experiments are then conducted on search space design, reward mechanism design, and trial budget. Results show that the framework can reliably execute, parse, and evaluate parameter search processes while preserving valid intermediate physical signals in complex designs. Experiments on AES further indicate that the global search space is more stable across random seeds, the narrow search space achieves better final solutions when successful, the stage-aware reward performs better for some seeds but is seed-sensitive, and increasing the trial budget from 10 to 20 helps markedly, whereas gains from 20 to 30 are limited.