(475d) Data Clustering-Based Hybrid-Kinetic Model of an Industrial-Scale Biochemical Fermenter | AIChE

(475d) Data Clustering-Based Hybrid-Kinetic Model of an Industrial-Scale Biochemical Fermenter


Shah, P. - Presenter, Texas A&M University
Sheriff, M. Z., Purdue University
Bangi, M. S. F., Texas A&M University
Kwon, J., Texas A&M University
Kravaris, C., Texas A&M University
With rapidly progressing technology and increasing energy demands, there is a growing need for cheap, environment-friendly energy sources. One of the promising candidates that fit these criteria is ethanol. Current ethanol production methods include chemical conversion of food sources like corn [1] and thermochemical conversion of lignocellulosic biomass [2]. However, they suffer from drawbacks like sustainability issues in both the food and fuel market and insufficiency in meeting the high energy demands. Hence, there has been a shift of focus towards biochemical fermentation, which can utilize different cellulosic biomass substrates and pretreatment technology options without using food resources, thereby solving the previous methods' downsides [3]. However, building a first-principles model is difficult as it requires extensive knowledge of reaction kinetics, thermodynamics, transport, and physical properties. These are difficult to obtain because of the intrinsic time-varying behavior of cell metabolism, thus limiting the prediction capability. [4]. To solve this challenge, data-based models such as neural networks have gained traction in this area of research, but they are data-dependent and show poor accuracy when tested outside the training regimes. To overcome both these modeling techniques' limitations, a hybrid-kinetic model framework was developed. This model's foundation is based on first principles, and the unknown properties of the underlying process are modeled using machine learning methodology.

The hybrid model is a combination of a first-principles model and a data-based model [5]. Due to uncertainties in the parameters and lack of complete process information, it is challenging to build a process model that can accurately predict the states resulting in a plant-model mismatch between predicted and measured state concentrations. In order to improve the accuracy of the model, global sensitivity analysis was performed by varying the nominal parameters in a specified range. This was done to identify the most critical parameters affecting the model prediction. Based on the knowledge of these sensitive parameters, the prediction horizon is divided into multiple time-intervals, where each time-interval consists of a data cluster, and the sensitive parameters are estimated separately in each cluster. Hence, instead of using just one estimate of parameters for the entire duration of the process, our approach uses time-varying parameter estimates.

Furthermore, to quantitatively approximate the variation of these parameters, a neural network-based approach was adopted. The parameters that varied the most based on the clustering results were chosen as the output of the neural network, and the model states like biomass, substrate, and product were chosen as inputs to the neural network. The parameters estimated using the neural network were then used in the first principles model, and the output concentrations were determined. The Levenberg Marquardt algorithm was used to train the neural network and iteratively update the weights and biases [6]. These updated network parameters are used in the next iteration, giving a new set of outputs. The process is repeated till the sum squared error of the outputs is less than a tolerance value.

As a case study, we applied this framework to a biochemical fermentation process. The results show a better estimation of uncertain time-varying parameters, thus considerably improving the prediction of output states over previous models. The hybrid-kinetic model was validated using different experimental data showing reasonably good agreement to the online measurements. To further improve the prediction power, data from multiple fermentation batches were used to train the network. To conclude, the hybrid kinetic model reduced the plant model mismatch and provided a better prediction accuracy and extrapolation capability for sets of process data.


  1. Thompson, Paul B. "The agricultural ethics of biofuels: the food vs. fuel debate." Agriculture 2.4 (2012): 339-358.
  2. Niziolek, Alexander M., et al. "Coal and biomass to liquid transportation fuels: process synthesis and global optimization strategies." Industrial & Engineering Chemistry Research 53.44 (2014): 17002-17025.
  3. Raftery, Jonathan Patrick. Continuous Biochemical Processing: Investigating Novel Strategies to Produce Sustainable Fuels and Pharmaceuticals. Diss. 2017.
  4. De Azevedo, S. Feyo, B. Dahm, and F. R. Oliveira. "Hybrid modelling of biochemical processes: A comparison with the conventional approach." Computers & chemical engineering 21 (1997): S751-S756.
  5. Galvanauskas, Vytautas, Rimvydas Simutis, and Donatas Levisauskas. "Application of evolutionary computing for hybrid model based optimization of biochemical processes." WSEAS Transactions on Information Science and Applications 2.5 (2005): 507-511.
  6. Bangi, Mohammed Saad Faizan, and Joseph Sang-Il Kwon. "Deep hybrid modeling of chemical process: Application to hydraulic fracturing." Computers & Chemical Engineering 134 (2020): 106696.