Deep learning‐based surrogate modeling and optimization for microalgal biofuel production and photobioreactor design

Abstract

Identifying optimal photobioreactor configurations and process operating conditions is critical to industrialize microalgae‐derived biorenewables. Traditionally, this was addressed by testing numerous design scenarios from integrated physical models coupling computational fluid dynamics and kinetic modeling. However, this approach presents computational intractability and numerical instabilities when simulating large‐scale systems, causing time‐intensive computing efforts and infeasibility in mathematical optimization. Therefore, we propose an innovative data‐driven surrogate modeling framework, which considerably reduces computing time from months to days by exploiting state‐of‐the‐art deep learning technology. The framework built upon a few simulated results from the physical model to learn the sophisticated hydrodynamic and biochemical kinetic mechanisms; then adopts a hybrid stochastic optimization algorithm to explore untested processes and find optimal solutions. Through verification, this framework was demonstrated to have comparable accuracy to the physical model. Moreover, multi‐objective optimization was incorporated to generate a Pareto‐frontier for decision‐making, advancing its applications in complex biosystems modeling and optimization. © 2018 American Institute of Chemical Engineers AIChE J, 65: 915–923, 2019

Hybrid physics‐based and data‐driven modeling for bioprocess online simulation and optimization

“…The length of the control (and prediction) horizon in this study was fixed to 2 days, similar to the industrial setup, thus the framework estimated eight control actions during each iteration. A hybrid stochastic optimization algorithm was designed to optimize the ANN chain, where random search was initially executed to narrow down the solution space and then simulated annealing was applied to refine the optimal solution (Ehecatl Antonio del Rio‐Chanona et al, ). This was executed in Mathematica 11.…”

“…Given the large size of accumulated data, models based on machine learning, particularly artificial neural networks, have been adopted to simulate bacterial fermentation and algal photo‐production (del Rio‐Chanona et al, ; Dineshkumar et al, ). Other advanced neural networks such as recurrent and convolutional neural networks have also been used to simulate the production of different biochemical (Valdez‐Castro et al, ; del Rio‐Chanona et al, ). Moreover, through the use of stochastic optimization, data‐driven models have been adopted to optimize long‐term fed‐batch processes, causing considerable increases in production of different metabolites and yielding highest intracellular contents of bioproducts reported to date (del Rio‐Chanona, Manirafasha, Zhang, Yue, & Jing, ; Dineshkumar et al, ).…”

Comparison of physics‐based and data‐driven modelling techniques for dynamic optimisation of fed‐batch bioprocesses

“…However, it is well known that neural networks present very many saddle points or local minima (Goodfellow, Bengio, & Courville, ), and therefore it is ill advised to use gradient‐based optimisation algorithms directly. As ANNs are very rapid at inference, the current study applied a hybrid evolutionary algorithm (del Rio‐Chanona et al, ; Estrada‐Wiese, del Río‐Chanona, & del Río, ) to initially optimise the ANN process model. Once the neighbourhood of a solution was found, a gradient‐based optimisation method was used to find the optimal inflow rate and light intensity.…”

Reinforcement learning for batch bioprocess optimization

“…This selection increases the log likelihood of an action by comparing it to the expected reward of the current policy. (16) is the gradient that we can now incorporate into our steepest ascent strategy. The algorithm that trains the RNN and obtains the optimal policy network is the following.…”

“…Bioprocesses exploit microorganisms to synthesize platform chemicals and high-value products by utilizing different means of resources [22]. Compared to a traditional chemical process, a biochemical process is highly complex due to the intricate relationships between metabolic reaction networks and culture fluid dynamics [16]. As a result, it is difficult to construct accurate physics-based models to simulate general large-scale biosystems, and plant-model mismatch is inevitable.…”