Combining Fundamental Models with Data

点击翻译按钮获取中文摘要

New methods will be developed for combining fundamental models and data, leading to accurate chemical-process models with reduced experimental costs and shorter development times. In the longer-term, the research will contribute to development of digital twins as advanced tools for industrial operation. Students will receive valuable training in modeling and data analysis, preparing them for careers that contribute to process digitalization and development of greener processes. The proposed research focuses on two persistent gaps that make it difficult to develop effective fundamental models of chemical and biochemical processes. The first is that it is difficult to use prior knowledge to select optimal experimental conditions that result in reliable parameter estimates and accurate model predictions. Although routine process data may be freely available, informative new experiments required for fundamental-model building can be expensive. Companies need better tools to plan experiments so that accurate models can be built in a shorter time with less cost. The second is is that it is difficult to tune dynamic models for use in on-line monitoring and control. State estimators are available for updating states and parameters in models as new data become available. However, tuning of state estimators (based on measurement and model uncertainties) remains a difficult problem. We will address the first gap by developing improved sequential methods for Model-Based Design of Experiments (MBDOE). The new methods will permit MBDOE calculations even when limited data (or a large number of model parameters) results in a non-invertible Fisher information matrix. The non-invertibility problem will be addressed using Bayesian terms in the MBDOE objective functions. The proposed approach will be compared with current methods using case studies on i) production of bio-sourced polyethers, ii) CO2 hydrogenation for renewable fuels production and iii) acrylate/methacrylate copolymerization for low-solvent automotive coatings. We will address the second gap using new simultaneous methods to estimate model parameters and obtain uncertainty information for state-estimator tuning. Fundamental models will be augmented with new empirical stochastic terms and state equations to account for disturbances and model imperfections. Fundamental model parameters, model uncertainties and measurement-noise variances will be estimated from old batches of dynamic data using efficient maximum-likelihood and Bayesian algorithms. The proposed simultaneous method will be tested against existing techniques for tuning extended Kalman filters (EKFs) and related estimators wherein model parameters and noise covariances are estimated in separate steps. We hypothesize that more-reliable parameter and state estimates will result from the proposed simultaneous approach, leading to improved on-line model predictions. An industrial polyethylene reactor model will be used as a case study.