Abstract:
Chemical feedstocks are complex and often heterogeneous and distributed. In the past, this was often addressed by suitable pretreatments or separation into pure feeds before processing. However, considerations of sustainability and a circular economy (which can, for example, lead to more recycling of products into feed streams) and investigating new feedstocks, such as renewables, have led to new challenges in processing feedstocks. Often, product specifications around selectivity and/or yields have also become more stringent. Efforts around process intensification can also create complexity. All these aspects pose significant challenges related to understanding reactions and processes choices for these systems.
Methods for understanding these systems range from empirical investigations to first principles (i.e., ab initio chemistry) investigations. Typical objectives in studying these systems range from understanding chemistry to developing reactor and process designs to monitoring and controlling processes. Typically, one would use empirical data (e.g., from spectroscopic sensors) to identify species, reactions and the progress of reactions, whereas ab initio methods would be used to understand specific reactions with well-specified reactants.
In this presentation, I will describe the development and performance of advanced models that leverage domain knowledge, machine learning and the principles of process systems engineering to address the challenges associated with both classes of problems: starting from on-line sensor data or from ab initio simulations. In the first case, the sensors can potentially track what is happening in the system, but the complexity of the reactions and the lack of knowledge of even the species that are present is a formidable challenge. In the ab initio approach, the challenge is posed by the fact that condensed phase simulations are difficult to perform (because of computational cost and complexity), but are essential to understanding the solvent mediated processes that are key to the processing of many renewable feedstocks.
The approach to developing reaction networks from spectroscopic sensor data includes using data fusion, joint non-negative tensorial factorization and Bayesian networks to identify pseudo-components and reaction networks. Convolutional neural networks are used to identify functional groups and generate candidate molecules from substructures, which are then matched to molecules present in chemistry databases. This is then followed by matching the pseudo-reaction networks to reaction templates in the databases to identify the dominant real chemistry. In addition, neural ODEs are used to identify mechanism-constrained kinetics, and hidden semi-Markov models are used to monitor changes in the dominant chemistry. Finally, model predictive controllers are developed based on these interpretable machine learning models.
For ab initio simulations, we have developed ML methods based on classification and the Mahalanobis distance to identify, from short bursts of simulation, whether solvent reorganization affects reaction energetics significantly. If these affects are significant, we have used LSTMs to build proxy models that can reduce the computational cost of simulations greatly.
Bio:
Vinay Prasad is the Jaffer Professor in Process Systems and Control Engineering in the Department of Chemical and Materials Engineering at the University of Alberta. He has degrees from the Indian Institute of Technology (IIT) Bombay, Kansas State University and Rensselaer Polytechnic Institute, all in chemical engineering. He has academic and industrial experience in three countries (India, the USA and Canada) across two continents that spans a variety of institutions.
His research has been characterized by a combination of theoretical innovations and real-world application. His research interests are in the application of hybrid and interpretable machine learning and first-principles approaches for discovery, modeling, estimation, monitoring, control and optimization. Significant research contributions include the development of data fusion methods and network determination for complex reacting systems, constrained estimation and soft sensing methods, and hybrid and interpretable machine learning for complex chemical processes. The real-world impact of this work has been seen in diverse areas such as modeling and optimization of carbon capture and storage systems, chemoinformatics-based discovery and monitoring of biomass conversion, monitoring and optimization of reservoirs, and soft sensing, estimation and control of mineral processing systems.