Optimization of process parameters in semiconductor annealing considering existing conditions

Process informatics has facilitated the application of artificial intelligence to optimize manufacturing conditions across various domains. However, when established process conditions already exist, AI proposed conditions that diverge significantly can be difficult to adopt. In this study, we introduce additional objective functions that measure the distance between candidate conditions and existing ones, guiding the search toward feasible solutions. We focus on a five-step heat treatment of silicon wafers by constructing a surrogate model based on a simulator of bulk microdefect (BMD) formation, growth, and dissolution. Using NSGA-II, we conducted a multiobjective optimization across 22 process parameters with a performance metric and three distance metrics: L1 norm, L2 norm, and count of matched conditions relative to the existing. The Pareto fronts obtained with distance metrics closely approximated those from conventional optimization without distance metrics, while yielding solutions nearer to existing conditions. Further analysis of optimal parameters revealed that L1 norm minimization produced sparse solutions identical to the baseline for some variables, whereas L2 norm minimization achieved consistently small deviations across all parameters. These findings demonstrate that incorporating norm-based distance metrics enables tailored selection of solutions that balance performance and transition feasibility.