Trace clustering is a key technique in process mining for reducing model complexity by grouping similar traces before process discovery. However, selecting appropriate combinations of feature encoding methods and clustering algorithms is often manual and lacks behavioral validation. This paper proposes an automated framework that systematically explores and evaluates combinations of five encoding methods and three clustering algorithms based on both clustering cohesion and process model quality. Unlike existing approaches that rely solely on structural metrics, our framework integrates process oriented evaluation criteria including fitness, precision, generalization, and simplicity metrics to select optimal configurations using a multi objective scoring and Pareto optimization approach. In addition, we incorporate a learning-based prediction mechanism using a neural network model to automatically classify and prioritize high-quality clustering configurations, further improving the scalability and generalizability of the framework across diverse event logs. Experiments on real world datasets, including the BPI Challenge 2020 and Road Traffic Fine Management logs, show that combinations such as DBSCAN with generalized edit distance consistently produce behaviorally coherent clusters and more interpretable process models. Experimental results show that considering downstream process model quality during clustering selection significantly improves the utility of trace clustering. The proposed framework offers a practical and extensible solution for automating trace clustering, reducing manual effort, and enhancing the effectiveness of process mining applications.