Abstract
Reliable process operation under uncertainty remains a fundamental challenge in chemical and pharmaceutical manufacturing. Variability arising from feed fluctuations, material properties, external disturbances, and uncertain model parameters can significantly impact feasibility and performance when operating ranges are designed under nominal assumptions. Design Space Identification (DSI) addresses these limitations by defining a Probabilistic Design Space (PDS), within which process constraints are satisfied with a prescribed confidence level. However, identifying and extracting robust and practically usable nominal operating ranges (NORs) from complex, high-dimensional, and nonconvex PDSs remains computationally demanding. This work proposes a novel reinforcement learning (RL)-based framework for automated identification of the largest robust hyper-rectangular NOR fully contained within a given PDS. The design centering problem is formulated as a sequential decision-making task and solved using the Proximal Policy Optimization (PPO) algorithm. By interacting with a probabilistic feasibility environment, the RL agent adaptively adjusts the location and size of the NOR to maximize its area while enforcing probabilistic constraint satisfaction. Deep neural network policies enable efficient exploration of nonlinear feasibility boundaries without exhaustive sampling or gradient-based optimization. The approach is demonstrated on a tablet lubrication case study with parametric uncertainty. Results show rapid convergence to well-located, near-maximal NORs using substantially fewer probabilistic feasibility evaluations than conventional grid-based or Monte Carlo methods. The proposed PPO-based framework provides a scalable and data-efficient solution for robust operating range identification in uncertainty-aware process design.