(288a) Bayesian Optimization of Crude Sulphate Turpentine Conversion to p-Cymene
AIChE Annual Meeting
2020
2020 Virtual AIChE Annual Meeting
Process Development Division
Big Data, Artificial Intelligence, and Machine Learning Examples for Process Research
Wednesday, November 18, 2020 - 8:00am to 8:15am
The identification of novel routes to synthetize functional molecules from bio-waste feedstock is one of the open challenges for a sustainable development of the chemical industry (Helmadach et al., 2017). However, most of the available feedstocks are complex mixtures of different molecules, which would ideally be processed without tedious and costly purification. The chemical complexity of such mixtures and their chemical transformations makes it difficult to build accurate mechanistic models for the optimization of the process conditions. To efficiently tackle this problem, we combine recent developments in machine learning (ML) (Mateos et al., 2019) and closed-loop optimization. The use of data-driven surrogate models coupled with design of experiments (DoE) algorithms have the potential to significantly speed up optimization of processes and suggest optimal conditions for synthetic routes, when no physical models are available.
The case study under consideration is the conversion of crude sulphate turpentine (CST), a complex mixture of organic molecules and isomers resulting from pulp and paper industry, to p-cymene, a highly-valuable chemical. The product p-cymene shows a wide range of antioxidant and biological activity, which makes it particularly relevant for the food and pharmaceutical industry. Further applications include the synthesis of fragrances, the use as a solvent for dyes and varnishes, and, it has been proposed as an alternative precursor of terephtalic acid(Eggersdorfer, 2003).
The proposed synthetic route consists of an acid-catalysed ring opening reaction to convert CST to a mixture of terpinene isomers followed by an aerobic dehydrogenation to p-cymene. The two steps of reaction are summarized in Scheme 1. Both steps suffer from selectivity issues. In the first reaction step, the acid catalyses polymerization reactions with an overall decrease of the overall yield to the products of interest. Unidentified by-products are also formed during the aerobic hydrogenation reaction.
The overall goal of this study is to perform a multi-objective optimization for maximization of the overall yield and conversion of each step of reaction, due to the selectivity problems. The two reaction steps are optimized individually using the Thompson sampling efficient multi-optimization algorithm (TS-EMO) (Bradford et al., 2018). Herein, we extended the algorithm to account for linear constraints in the input space and the sampling time under batch conditions. The results highlight the great potential of closed-loop optimization for identifying optimal conditions in a complex high-dimensional space. Notably, superior operating conditions were found within a low number of experiments. Moreover, the results show that experimental errors and inaccurate measurements still pose a challenge to automated reaction optimisation that should be addressed in future research.
Main results and conclusions.
The first step of reaction was carried out in a stirred batch reactor under kinetic regime, in the presence of an immiscible sulfuric acid aqueous solution. A closed-loop optimization was carried out for the first step of reaction considering eight input variables: molar fractions of the compounds in the CST mixture, residence time, temperature, concentration of sulfuric acid in the aqueous phase, and ratio between the organic and the aqueous phase. Batches of 4 reactions were simultaneously carried out at each iteration, using a batch-sampling strategy. After 60 experiments, the model was able to suggest a dozen of consecutive optimal solutions, confirmed experimentally. Interestingly, the top 5 best solutions, with conversions higher than 95% and yield higher than 80%, were obtained under a wide variety of different conditions, suggesting the potentiality of the adopted algorithm to efficiently explore high-dimensional spaces without getting stuck in local optima.
The second reaction step was carried out feeding a gaseous stream of air and the liquid organic substrate to a continuous flow build-in reactor with recycle, packed with glass inert sphere to increase the contact surface between the phases (Plucinski et al., 2005), using tert-butyl hydroperoxide as a radical initiator and p-xylene as a solvent. Again, eight optimisation variables were considered as the ingredients of the organics in the liquid phase, the concentration of the radical initiator, temperature, liquid and gas flow rates, and reaction time. In this case, a single sampling strategy was applied, because only one reactor was available. Surprisingly, the best solutions were identified within the first 36 experiments (including the initial data set, 13 experiments). This can be mainly ascribed to two main factors: (i) the smaller experimental error associated to the GC-MS analyses, with respect to the NMR used in the first case; and (ii) the single sampling strategy is generally more efficient than the batch sampling.
The results show great potentiality of TS-EMO to deal with multi-optimization of high-dimensional problems in a relatively short time and reducing the amount of experimental resources, without any required mechanistic knowledge about the chemical system under investigation. Further research will be needed to improve the performances in the presence of significant experimental errors and the batch-sampling strategy and to efficiently integrate discrete variables (solvent selection, etc.) in the multi-optimization problem.
References
- Helmdach, P. Yaseneva, P.K. Heer, A.M. Schweidtmann, A.A. Lapkin, Chem. Sus. Chem. 2017, 10, 3632 â 3643.
- Mateos, M.J. Nieves-Ramacha, J.A. Rincon, React. Chem. Eng. 2019, 4, 1536 â 1544.
- Bradford, A.M. Schweidtmann, A.A. Lapkin, J. Glob. Optim. 2018, 71, 407 â 438.
- Eggersdorfer, Terpenes, Ullmannâs Encyclopedia of Industrial Chemistry, Wiley-VCH.
P.K. Plucinski, D.V. Bavykin, S.T. Kolaczkowski, A.A. Lapkin, Ind. Eng. Chem. Res. 2005, 44, 9683 â 9690.