Highly variant and immature production processes frequently occur during the ramp-up of new manufacturing technologies, where process parameters, resource configurations, and operating strategies must be established under incomplete process knowledge. The resulting combinatorial configuration spaces make systematic planning and evaluation of alternative process setups difficult. This paper proposes a planning-based digital twin architecture that enables structured exploration and reduction of such decision spaces while integrating simulation-based feasibility assessment. The approach combines formal modeling of discrete process parameter variants with automated planning and physics-based simulation using PyBullet. Symbolic planning operates on an abstract representation of process steps, resources, and parameter variants to generate consistent process paths under cost and quality objectives. These process paths are subsequently instantiated as executable simulation scenarios, allowing verification of their physical feasibility and operational behavior. The architecture integrates decision generation, execution, and evaluation within a unified and modular pipeline. A virtual thermoforming demonstrator is used to verify the functional feasibility of the approach and to illustrate systematic decision space reduction through constraint-based planning. The results demonstrate that planning-based digital twin architectures provide a scalable foundation for supporting decision-making during ramp-up and configuration of highly variant production processes.