Sustainable and resource-efficient production practices have motivated the development of advanced control and dynamical optimization strategies in process industries. Despite increasing attention to integrated industrial networks, there exists potential limitations on systematically applying data-driven ODE models and advanced control algorithms to establish circular economy (CE) synergy for waste minimization. Within a dynamically operating manufacturing network, waste circularization between the industrial systems under real operational constraints such as stock buffer capacities and commodity transport delays need to be effectively studied. We aim to demonstrate a combined hybrid approach leveraging Sparse Identification of Nonlinear Dynamics (SINDy) for model development and a receding-horizon Model Predictive Control (MPC) framework for solving the optimization in real-time. This work focuses on a pharmaceutical drug manufacturing network integratively functioning to produce the drug Acetaminophen (APAP), with the material exchange network comprising Acetic Anhydride (AA), Para-Aminophenol (PAP), Acetaminophen (APAP), Green Hydrogen (H2), and Urea Fertilizer (UF). Each plant generates a main product, byproducts, and waste commodities, except H2. The APAP production process requires both AA and PAP, which are delivered through buffer stocks situated at each upstream plant, subject to transportation delays. An APAP waste is circularized to the H2 process, while one PAP waste is circularized to the UF process. We aim to minimize the overall waste generation across the network while meeting the production demand of APAP, H2, and UF processes, while minimizing total non-circular waste. First, time-series data are collected from each industry using high-fidelity simulations and this data is used to recover parsimonious ODE-based models that capture the underlying nonlinear dynamics of each process using SINDy. These SINDy-identified equations are then integrated into an MPC scheme that explicitly accounts for transport lags and stock evolution. Such transportation delays are modeled using dead-time operators which shift the shipment flows forward in time, ensuring each plant only consumes materials that have truly reached its destination. Buffer stock balances are enforced through mass-balance constraints that can effectively prevent overflows and shortfalls, while the circularized waste commodities always get exchanged with the respective processes. Using a centralized MPC formulation, each prediction horizon update will yield an optimal set of control feed rates over discretized time steps, factoring in forecasted demands, stock capacities, and environmental limits on waste. The control loop repeats at regular intervals with each new iteration, incorporating real-time state measurements which will result in continual adjustments to the unanticipated disturbances such as production slowdowns or fluctuating transport times, thereby maintaining feasible and cost-effective operations in the entire manufacturing network. This approach will confirm that such an integrated, model-based approach can minimize the total waste while maintaining the exchange of dominant waste, promoting circular economy. Real-world practitioners can tailor the system by refining the objective function and constraints to comply with specific environmental regulations and to meet varying APAP, H2, and UF demand profiles, fostering sustainable and efficient operation within a complex circularizing pharmaceutical industrial network.
