Physics-informed digital twin design for supporting the selection of process settings in continuous manufacturing, with a focus in fiberboard production

Abstract

In process industry, plant operators often rely on their experience to choose suitable process settings that meet the productivity and quality goals. When these goals are not met, multiple changes to the settings might be necessary, which is time-consuming because each adjustment requires waiting for the new steady-state condition. A digital twin that quickly provides key performance indicators in steady-state as a function of these settings can speed up this task. The settings can be manually simulated before being adopted, or the digital twin can be integrated into an optimizer to automatically suggest optimal values to the operator, who ultimately makes the final decision. Despite advances in approaches to design such digital twins, most studies lack strategies to update the models when the plant behavior changes, and often overlook constraints and human-centric aspects of the plant operation. To address these gaps, we present a framework for training, tuning, and updating models for supporting the selection of process settings in continuous manufacturing. By directly mapping the steady-state conditions as a function of process settings, our approach enables informed decision-making and paves the way towards process optimization without requiring modifications to the plant control software, a crucial factor in established plants to ensure safety. We propose an interpretable model architecture, and a training process that incorporates both data and prior physical knowledge. Triggers detect deviations between the models’ predictions and the plant condition, in order to start model updates. The procedure for updating the models is tuned to perform consistently well in a variety of conditions, based on substantial simulations in historical data. To select the triggers, we balance technical and human aspects, by considering the trade-off between frequent model updates, increasing operator workload with frequent settings changes, versus how closely the models track the plant conditions. The framework is applied to five different stages of the fiberboard production process in a 1.4-year dataset, to predict key energy and quality-related variables as a function of process settings. The results show that the models, when connected to the data stream, are effectively updated when needed, show high sensitivity to the process settings and consistency with the available physical knowledge, making them well-suited to support the selection of process settings.