AI/ML in Process Systems Engineering
Recent improvements in state-of-the-art experimental and computational infrastructures, affordability, automation, ubiquitous connectivity through IoT, global push towards meeting environmental constraints to ensure safety and sustainability resulted into generation, processing and management of enormous amounts of heterogeneous data in the domain of Process Systems Engineering (PSE). PSE, which deals with the process design for the purpose of converting raw goods to usable end products, focuses on the design, operation, control, optimization and intensification of chemical, physical, and biological processes. Our aim @ GOKUL is to develop state-of-the-art data tools that can cater the special needs for vast amounts of highly complex data generated by the PSE community.
We target potential areas in PSE and investigate how the applications of deep supervised / unsupervised learning methods based on artificial neural networks (ANN) can be made useful there. Exploiting the novel multi-objective evolutionary Neural Architectural Search technique developed @GOKUL, TRANSFORM, we could successfully show how (i) optimization of computationally expensive models can be improved multiple folds using surrogate models (ANN), (ii) accurate system identification and data based model predictive control of extremely nonlinear industrial processes can be performed (RNN, LSTM), (iii) image based sensing can be improved for better optimization of the process (CNN, AE, VAE), (iv) the uncertainty quantifications for nonlinear models using analytical derivations can be obtained through Sobol indices and global sensitivity analysis (PUNNs), (v) the ideas of approximation of control vector using ANNs can be utilized to solve complex single and multi-objective optimal control problems efficiently, (vi) fuzzy clustering performance can be improved by neural networks based reformulation for identification of global optimum and (vii) generative modelling can be utilized to accurately solve the industrial nonlinear multi-objective optimization problems in uncertain framework (GAN, VAE). Our targeted applications include wind farm layout optimization, new alloy discovery by enhanced computational materials science calculations, monitoring environmental parameters due to climate change, smart sensing of particulate matter, fast charging protocols in Li+ battery management, bio-fuel supply chain optimization, systems biology (cell classification based on Ca+ oscillations in neurons), chemical engineering (polymerization reactors), metallurgical engineering (steel making processes etc.), mineral processing (grinding and flotation) and mechanical engineering (uncertainty analysis in supersonic flow of tactical missiles, surrogate optimization using CFD models) applications. Apart from the desired tangible benefits, some of these results brought laurels to GOKUL as they were bestowed with the best paper award in the international platforms (e.g. ACODS in 2020), best B. Tech project award by INAE (2021) and highlighted as works which can open up new opportunities to explore new designs in future (e.g. BATTERY 2030+, a long-term roadmap for forward-looking battery research in Europe, prepared by the EU Horizon 2020 initiative mentions our work in the Li+ Battery space). We acknowledge the collaborations with universities in USA (U Texas, Austin; U Washington, Seattle; Washington U, St Louis; U Miami), UK (U Exeter), Australia (Deakin U; Swinburne U) and funding agencies (UKIERI, MHRD, DBT, DST, DRDO, Tata) for their support.
Tadepalli, A., Pujari, K. N., Mitra, K., A Crystallization Case Study towards Optimization of Expensive to Evaluate Mathematical Models using Bayesian Approach, Materials and Manufacturing Processes, 2023, DOI: 10.1080/10426914.2023.2238051- Impact Factor 4.8.
Manoj, A., Miriyala, S. S., Mitra, K., Multi-objective Optimization Through a Novel Bayesian Approach for Industrial Manufacturing of Polyvinyl Acetate, Materials and Manufacturing Processes, 2023, DOI: 10.1080/10426914.2023.2195915 - Impact Factor 4.8.
Ravi Kiran, I., Naik, S., Mitra, K., Towards Faster Operational Optimization of Cascaded MSMPR Crystallizers using Multi-objective Support Vector Regression, Ind. Eng. Chem. Res. 2022, 61, 11518−11533 - Impact Factor 4.31.
Krishnan, K. J., Mitra, K., A Modified Kohonen Map Algorithm for Clustering Time Series Data, Expert Systems With Applications, accepted - Impact Factor 8.67.
Gumte, K., Pantula, P. D., Soumitri M. S., Mitra, K., Data Driven Robust Optimization for Handling Uncertainty in Supply Chain Planning Models, Chemical Engineering Science, 2021, 246, 116889 - Impact Factor 4.31.
Sharma, S., Pantula, P. D., Soumitri M. S., Mitra, K., A Novel Data-driven Sampling Strategy for Optimizing Industrial Grinding Operation under Uncertainty using Chance Constrained Programming, Powder Technology, 2021, 377, 913-923 - Impact Factor 5.13.
Ravi Kiran, I., Soumitri M. S., Mitra, K., Recurrent Neural Networks based Modelling of Industrial Grinding Operation, Chemical Engineering Science, 219, 115585 - Impact Factor 4.31.
Pantula, P. D., Mitra, K., Towards Efficient Robust Optimization using Data based Optimal Segmentation of Uncertain Space, Reliability Engineering and System Safety, 197, 106821 - Impact Factor 6.19.
Pantula, P. D., Soumitri M. S., Mitra, K., An Evolutionary Neuro-Fuzzy C-means Clustering Technique, Engineering Applications of Artificial Intelligence, 89, 103435 - Impact Factor 6.21.
Soumitri, M. S., Mitra, K., Deep Learning based System Identification of Industrial Integrated Grinding Circuits, Powder Technology, 360, 921 - 936 - Impact Factor 5.13.
Pantula, P. D., Mitra, K., A Data-Driven Approach Towards Finding Closer Estimates of Optimal Solutions Under Uncertainty for an Energy Efficient Steel Casting Process, Energy, 189, 116253 - Impact Factor 7.15.
Soumitri M. S., Subramanian, V., Mitra, K., TRANSFORM-ANN for Online Optimization of Complex Industrial Processes: Casting process as Case study, European Journal of Operational Research, 264, 2018, 294-309 - Impact Factor 5.33.
Virivinti, N., Mitra, K., Fuzzy Robust Optimization for Handling Feed Stream and Model Parameter Uncertainties during Comminution Process, Journal of the Taiwan Institute of Chemical Engineers, 70, 2017, 411-425 - Impact Factor 5.88.
Systems Biology
Subtilin production is favorable when Bacillus subtilis 168 is subjected to stress condition such as nutrient scarcity. A mathematical model underlying such growth process has immense applicability in determining the optimal operating conditions at industrial scale. We present this work with multiple objectives of a) selection of a substrate for creating the minimal nutrient media for B. subtilis thereby enhancing subtilin production, b) experimental study of the growth along with morphological characteristics of B. subtilis and product profile in nutrient scarcity condition and c) identification of an optimal unstructured model for subtilin production using a computational framework. First, we show that subtilin can be produced while B. subtilis is grown using galactose and B. subtilis undergoes morphological changes and takes filamentous shape. We then constructed a series of plausible models and used a hybrid method combining Genetic Algorithm and gradient based search methodologies, for model selection. The estimated kinetic parameters and the stoichiometric analysis indicate that the B. subtilis growth/death, product profile and respiratory mechanism undergo specific modifications in galactose as an adaptive response. Current study provides an inexpensive platform to produce subtilin and the predictive framework presented here has potential applications for large scale production of subtilin.
Greener process synthesis for drug production focuses on replacement of chemical process with biological processes. However, optimization of these processes remains challenging due to lack of experimental data particularly for long term passaging and predictive models that can accurately emulate the data. In this context we propose a scheme containing the co-culture of E. coli and B. subtilis that can be used to produce indole. The cell growth and product profiles obtained from fermenter corresponding to pure culture and individual cultures after passaging was used for unstructured model selection. Kinetic parameter estimation was performed using a hybrid technique combining classical and evolutionary algorithm. The major novelty of the work is the implementation of optimal control for dynamic parameter estimation to capture the interaction between two organisms in long term passaging. The model is capable of emulating the mixed culture effect on cell growth, product formation and substrate depletion. This is the first instance where an integrated framework combining experimental and computational methodology is used to predict the product profiling in mixed culture of E. coli and B. subtilis towards indole production. Such a framework can be used for optimizing the process parameters to increase the indole production at industrial scale.
Sharma, S., Mahadevan, J., Giri, L., Mitra, K., Identification of Optimal flow rate for Culture media, Cell Density and Oxygen towards maximization of Virus production in a fed-batch Baculovirus-Insect cell system, Biotechnology and Bioengineering, accepted- Impact Factor 3.8.
Dhyani V., George K., Gare S., Venkatesh K. V., Mitra K., and Giri L., A computational framework for estimation of neurotransmitter level and channel parameters from Ca2+ transients in dissociated hippocampal neurons, Hippocampus, 2023, accepted - Impact Factor 3.5.
Sharma, S., Sarkar, R., Giri, L., Mitra, K., Computational framework to understand the clinical stages of COVID-19 and visualization of time course for various treatment strategies, Biotechnology and Bioengineering, 2023, 120, 6, 1640-1656 - Impact Factor 3.8.
Sharma, S., Pujari, K. N., Giri, L., Mitra, K., Towards Performance Improvement of a Baculovirus-Insect Cell System Under Uncertain Environment: A Robust Multi-objective Dynamic Optimization Approach for Semi-batch suspension culture, Ind. Engg. Chem. Res., 2023, 62, 1, 111–125 - Impact Factor 4.32.
Kankanamge, D., Ubeysinghe, S., Tennakoon, M., Pantula, P. D., Mitra, K., Giri, L., Karunarathne, A., Dissociation of the G protein βγ from the Gq-PLCβ complex partially attenuates PIP2 hydrolysis, Journal of Biological Chemistry, 2021, 296, 100702 - Impact Factor 5.49.
Singh, R., Sharma, S., Kareenhalli, V. V., Giri, L., Mitra, K., Experimental investigation into Indole production using Passaging of E. coli and B. subtilis along with unstructured modeling and parameter estimation using Dynamic Optimization: An Integrated Framework, Biochemical Engineering Journal, 2020, 163, 107743 - Impact Factor 4.45.
Swain, S., Gupta, R. K., Ratnayake, K., Priyanka, P., Singh, R., Jana, S., Mitra, K., Karunarathne, A., Giri, L., Confocal imaging and k-means clustering of GABAB and mGluR mediated modulation of Ca2+ spiking in hippocampal neurons, ACS Chemical Neuroscience, 9 (12), 3094-3107 - Impact Factor 5.78.
Singh, R., Miriyala, S. S., Giri, L., Mitra, K., Kareenhalli, V. V., Identification of Unstructured model for Subtilin production through Bacillus subtilis using Hybrid Genetic Algoritm, Process BioChemistry, 60, 2017, 1-12 - Impact Factor 4.89.
Gupta, R. K., Swain, S., Kankanamge, D., Pantula D. P., Singh, R., Mitra, K., Karunarathne, A., Giri, L., Comparison of calcium dynamics and specific features for G-protein coupled receptor targeting drugs using live cell imaging and automated analysis, SLAS DISCOVERY: Advancing Life Sciences R&D, 2017, 22, 848-858 - Impact Factor 3.34.
Energy (Wind/Energy Storage/Bio-Energy/Climate Change)
Climate Change
Climate change is increasingly gaining central stage in the discussions pertaining to public health and national policy making. To facilitate informed decision making for the policy makers, it is necessary to provide predictive models which accurately represent the amount of Green House Gas (GHG) and Particulate Matters (PM) like PM_10,PM_2.5,CO,CO_2,SO,SO_2,CH_4, pH of rainfall etc. in environment The availability of vast amounts of data from sensors is driving the use of AI/ML based techniques for understanding climate change phenomenon. Researchers at Global Optimization and Knowledge Unearthing Lab (GOKUL) have designed optimal Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks to capture the dynamic trends of ~15 environmental parameters from the data that cause long-term health hazards. Complicated cross-correlations among various GHGs, PMs and pH of rain were established by performing Global Sensitivity Analysis (GSA), which helped in building the most effective functional mapping among influential features and the output variables through multi-variate modelling.
Loss of Crop, unless predicted and corrected ahead in time, can create a havoc in society’s exchequer every year. To address this issue, models which can identify the irregular or abnormal growth of the crop would be of great help. The satellite imagery data of the crops throughout the entire year can be used to build such models. GOKUL’s research is currently focusing on designing a predictive framework, which can capture the growth pattern of different crops from these satellite images. The model includes a combination of supervised and unsupervised learning methods to extract the spatial and temporal dynamics from the images. This model thus learns the growth pattern of the crops at every time step and identifies any abnormal growth indication at any time throughout their growth period. The causes of these abnormalities (such as drought conditions, weather changes, soil contamination etc.) can be found out and necessary measures to address these causes can be taken at early stages leading to prevention of crop loss and enhancement of yield.
Wind Energy
Even-though state-of-the-art technologies are developed across the world to harness the renewable energy, the efficiency remains low due to the uncertain and nonlinear nature of such resources. Though India is abundantly blessed with a potential wind resource, large-scale profitable establishment for renewable energy conversion systems in India is rarely seen due to the uncertainty associated with it. Additionally, a common problem faced in the domain of windfarm modelling is the computational expense related to simulating the entire study. Thus, windfarm layout optimization, wake modelling, uncertainty handling and control studies during energy harnessing from wind are still at inception with-respect-to Indian subcontinent.
Development of robust wind energy conversion system is need-of-the-hour to offset the energy crisis and drastic environmental issues India is facing in current times. At Global Optimization & Knowledge Unearthing Lab (GOKUL), we proposed novel methodologies to design optimal windfarms from grassroots level by combining the fields of deep learning, CFD, combinatorial & evolutionary optimization and uncertainty analysis. Further, we are working on new robust windfarm control strategies using reinforcement learning. Such a unique framework can resolve several issues faced by wind farm owners and ensure designs that can last for long-term duration. The current status, scope for improvement and novelties in the proposed work are presented below:
• Realizing the need for forecasting wind characteristics data due to their limited availability, a novel heuristic-free optimal design algorithm for building nonlinear deep learning-based system-identification techniques has been proposed. Also, the abilities of generative models like Gaussian Processes and Variational Autoencoders are utilized in combination with a clustering based generative model for accurately modelling the uncertain nature of wind.
• Methods for modelling turbulent wake effects in wind farms are focused on next. Here, machine learning driven accurate models are developed, which would be fast as compared to high fidelity CFD based models.
• As a whole, the windfarm layout optimization problem turns out to be NP-hard MINLP formulation. To convert such a large-scale problem to small scale, an auto-encoder based strategy is proposed, which assists in efficient usage of combinatorial, evolutionary and hybrid optimization algorithms for micrositing.
• To make it realistic, wind state uncertainty in the micrositing formulation is considered and solved using Robust Optimization. Moreover, wind farm control studies using reinforcement learning are performed.
The novel ideas developed in the lab have led to high-impact peer-reviewed publications of international repute and two high-value Government projects worth INR 1 crore with international collaboration for establishment of large-scale efficient windfarms. The funding agencies include DST - National Supercomputing Mission, SPARC – MHRD and special fund by British Council UKIERI through the international collaboration with University of Exeter, UK (Department of Computer Science).
Bio-Energy Supply Chain
One of the vibrant global issues i.e. how to tackle the increase in energy demand with depleting conventional fossil fuels is currently addressed through optimal usage of non-conventional energy sources namely solar, wind, tidal, geothermal and bio sources. For a developing nation like India, where 70% of population depends on forest and agriculture, the bioenergy sector can play a vital role that is yet to be utilized to its full potential. The national initiatives towards blending 20% biofuels with fossil fuels is catalyzing this fact further. Some of the existing challenges in this sector are (i) biofuels have calorific value less than fossil fuels leading to more quantity of biomass needed for energy generation, (ii) need of flexi-fuel engines costlier than regular engines, (iii) food vs fuel issues for 1st generation biomass, (iv) availability of biomass throughout the year due to seasonal nature of crops in addition to their heterogeneity in composition across geographies. Despite a lot of research happening at individual levels to ameliorate the current state-of-the-art technologies for energy conversion from different bio resources, a novel approach has been adopted by us to attack these problems holistically from the vantage point of a supply chain (SC) network designer.
A SC bridges several entities present in different echelons (material supply, manufacturing, distribution and collection) involved in converting the raw material into the finished goods and enables a designer to find possible avenues of improvement in the whole product life cycle. An endeavor towards designing such a country wide supply chain network has been successfully attempted for the first time considering the target of blending 20% of both bioethanol and biodiesel for a future time horizon (2018-2026) using 2nd generation biomass. The objectives of the SC to maximize the profit net present value and simultaneously minimize the pollution causing greenhouse gas emissions (GHGe) have been achieved through mixed integer linear programming, which is NP hard to solve. External importers are also included to manage short fall in indigenous production and to maintain the product quality in terms of research octane number. The unique SC model covers all three aspects of technology, economy and environment keeping sustainability in mind. From technology side, the model deals with choice of site location, capacity planning, multi-connection routing, choice for mode of transport considering biomass to biofuel conversion yields for several raw material choices to handle seasonality issues. Considering the time value of money and depreciation, economic calculations are performed not only tackling the capital and operating expenditure, but also through the GHGe emission, GHGe savings and conversion of carbon savings into carbon credits representing the environmental aspects. Further, to make the SC design realistic, stochasticity in biofuel demand, import price and biomass feed supply has been modeled using data driven robust optimization approach. Overcoming the drawbacks of conventional robust optimization, the adopted approach performs accurate transcription of uncertain parameter space using unsupervised machine learning approaches, which resulted in more accurate, non-conservative robust solutions. In addition to bestowed with the best paper award by International Federation of Automatic Control conference (ACODS 2020), the project findings are published in the prestigious International Journal of Cleaner Production on several occasions.
Pujari K, Miriyala, S. S. Mittal, P., Mitra, K., Better Wind forecasting using Evolutionary Neural Architecture Search driven Green Deep Learning, Expert Systems With Applications, 214, 2023, 119063 - Impact Factor 8.67.
Ravi Kiran, I., Soumitri M. S., Mitra, K., Deep Learning Based Dynamic Behaviour Modelling and Prediction of Particulate Matter in Air, Chemical Engineering Journal, 2021, 426, 131221 - Impact Factor 13.27.
Gumte, K., Pantula, P. D., Soumitri M. S., Mitra, K., Achieving Wealth from Bio-Waste in a Nationwide Supply Chain Setup under Uncertain Environment through Data Driven Robust Optimization Approach, Journal of Cleaner Production, 2021, 291, 125702 - Impact Factor 9.29.
Mittal, P., Mitra, K., In Search of Flexible and Robust Wind Farm Layouts Considering Wind State Uncertainty, Journal of Cleaner Production, 248, 119195 - Impact Factor 9.29.
Gumte, K., Mitra, K., Bio-Supply Chain Network Design to tackle ethanol deficiency in India: A mathematical framework, Journal of Cleaner Production, 234, 2019, 208-224 - Impact Factor 9.29.
Dawson-Elli, N., Kolluri, S. K., Mitra, K., Subramanian, V., On The Creation of a Chess-AI-Inspired Problem-Specific Optimizer for the Pseudo Two-Dimensional Battery Model Using Neural Networks, J Electro Chemical Society, 166, 2019, A1-A15 - Impact Factor 4.32.
Mittal, P., Mitra, K., Determining Layout of a Wind Farm With Optimal Number of Turbines: A Decomposition Based Approach, Journal of Cleaner Production, 202, 2018, 342-359 - Impact Factor 9.29.
Dawson-Elli, N., Lee, S. B., Pathak, M., Mitra, K., Subramanian, V., Data Science Approaches for Electrochemical Engineers - An Introduction through Surrogate Model Development for Lithium-Ion Batteries, J Electro Chemical Society, 165, 2018, A1-A15 - Impact Factor 4.32.
Mittal, P., Mitra, K., Kulkarni, K., Optimizing the number and locations of turbines in a wind farm addressing energy - noise trade-off: A hybrid Approach, Energy Conversion and Management, 132C, 2017, 147-160 - Impact Factor 9.7.
Mittal, P., Kulkarni, K., Mitra, K., A Novel Hybrid Optimization Methodology to Optimize the Total Number and Placement of Wind Turbines, Renewable Energy, 86, 2016, 133–147 - Impact Factor 8.0.
Optimization of Polymerization Reactors
Satisfaction of various objectives such as maximization of number average molecular weight along with minimization of Poly Dispersity Index, minimization of batch time do not necessarily guarantee the maximization of concentration of desired species during polymerization process. As the final product consists of a number of polymer species, a need is felt to perform an advanced optimization study to come up with such process conditions for which the selective growth of a particular polymer species is maximized in minimum possible processing time and the population of other species should be at their lowest values. These above-mentioned conflicting objectives frame the platform for a multi-objective optimization problem, which is solved here using a real-coded non-dominated sorting genetic algorithm or NSGA II and Pareto optimal solutions are obtained for a semi-batch epoxy polymerization process. The decision variables are discrete addition rates of various ingredients, e.g. the amount of addition of bisphenol-A (a monomer), sodium hydroxide and epichlorohydrin at different time steps. All species balance equations, bounds on Mn, PDI and addition amounts are treated as constraints. Results are very promising in terms of optimized operations for selective enhancement of desired polymer species for the epoxy polymerization process. Total additions are kept very close to available experimental conditions to minimize probable extrapolation errors. It has been observed that preferential oligomer production is extremely difficult for epoxy polymerization. Lower chain polymers are the only choice for a good quality, stable polymer product. Similar studies are performed for other polymerization systems such as PPT system, long chain branching of PVAc and PP systems with different objectives in mind.
Mitra, K., Genetic Algorithms in Polymeric Material Production, Design, Processing and Other Applications, International Materials Review, 53(5), 2008, 275-297 - Impact Factor 19.56.
Mogilicharla, A., Mitra, K., Majumdar, S., Modeling of propylene polymerization with long chain branching, Chemical Engineering Journal, 246, 2014, 175-183 - Impact Factor 13.27.
Mitra, K., Assessing Optimal Growth of Desired Species in Epoxy Polymerization under Uncertainty, Chemical Engineering Journal, 162, 2010, Page 322-330 - Impact Factor 13.27.
Majumdar, S., Mitra, K. and Raha, S. Optimized Species Growth in Epoxy Polymerization with Real Coded NSGA II, Polymer, 46, 2005, Page 11858-11869 - Impact Factor 4.43.
Majumdar, S. and Mitra, K., Modeling of a Reaction Network and its Optimization by Genetic Algorithm: A Novel Approach, Chemical Engineering Journal, 100, 2004, Page 109-118 - Impact Factor 13.27.