Abstract
The reliable identification of feasible and optimal operating conditions is a key challenge in the design and optimization of thermochemical conversion processes, where kinetics, limited data availability, and strict physical constraints coexist. In this work, a novel data-driven strategy based on Physics-Informed Neural Networks (PINNs) is proposed to explore the operability space of a bubbling fluidized bed (BFB) plastic pyrolysis process. The approach integrates mechanistic knowledge through explicit mass balance constraints with data-driven learning, enabling accurate prediction of and feasibility boundaries. An adaptive sampling framework is employed to iteratively augment the training dataset. The trained PINN surrogate is then used to predict feasible regions and perform constrained optimization aimed at minimizing tar production, which is one of the most problematic byproducts in plastic pyrolysis processes. Beyond classical optimality, a robustness-oriented uncertainty quantification methodology is introduced, combining local perturbations with feasibility filtering to assess the reliability of the optimal solution. Quantitative robustness metrics, including relative objective uncertainty, feasibility retention ratio, and an overall robustness index, are introduced and used to track convergence and select the most reliable operating point. Results show that the proposed framework efficiently converges to accurate feasibility regions and identifies operating conditions that balance optimal performance with robustness against uncertainty. This work provides a methodology for physics-driven process optimization and operability analysis, with potential applicability to a wide range of complex energy and chemical systems.