(504e) Machine Learning in Food and Drink Manufacturing
AIChE Annual Meeting
2024
2024 AIChE Annual Meeting
Topical Conference: Next-Gen Manufacturing
Next-Gen Manufacturing in Pharma, Food, and Bioprocessing
Wednesday, October 30, 2024 - 9:16am to 9:35am
Food and beverage industries receive key feedstocks whose composition is subject to geographic and seasonal variability, and rely on factories whose process conditions have limited manipulation margins but must rightfully meet stringent product quality specifications. Unlike chemicals, most of our favourite foods and beverages are highly sensitive and perishable, with relatively small profit margins. Although manufacturing processes (recipes) have been perfected over centuries or even millennia, quantitative understanding is limited. Predictions about the influence of input (feedstock) composition and manufacturing (process) conditions on final food/drink product quality are hazardous, if not impossible, because small changes can result in extreme variations. A slightly warmer fermentation renders beer undrinkable; an imbalance among sugar, lipid (fat) and protein can make chocolate unstable.
Though Food Science cannot provide global composition-structure-quality correlations, Artificial Intelligence/AI can be used to extract valuable process knowledge from factory data. For example, food colloid/pulp/paste thickness (viscosity) is instrumental in computational design and optimisation of key processing units (mixers, extruders): it varies enormously for small composition/temperature changes, but must be accurately estimated.
The case of beer, in particular, has been the focus of several of our papers in the past 8 years (2016-24), offering a sound comparison basis for evaluating model fidelity between published precedents and new PINN approaches. Pursuing PINN modelling can cater to greater complexity, both in terms of plant flowsheet as well as target product structure and chemistry, because ML/PINN tools can efficiently predict complex rheological behaviour (food colloids/pastes), which is itself instrumental in computational design and optimisation of key food processing units (mixers, extruders). Traditional (first-principles) descriptions of these necessitate elaborate, high-fidelity (e.g. CFD) submodels of extreme complexity, with at least two severe drawbacks: (1) cumbersome prerequisite parameter estimation with extreme uncertainty, (2) prohibitively high CPU cost.
The currently mature plantwide modelling technology faces dire challenges for industrial implementation in the Food+Drink Manufacturing sector, a lot more so than in other domains (e.g. fuels/pharmaceuticals), due to the ubiquitous multiphase and multicomponent mixtures. Transport properties (e.g. viscosity) vary over orders of magnitude versus composition and temperature. Such extreme variations perplex design and optimisation of manufacturing processes, by inducing structural and parametric uncertainty in first-principle models. Data-driven modelling, however, requires prohibitively expensive plant-wide instrumentation (the norm in refineries, but not at all in food+drink factories, with profit margins significantly lower).
Artificial Neural Networks/ANN and their representational versatility for process systems studies is known for decades [2]. First-principles knowledge, though (mass-heat-momentum conservation, chemical reactions) is captured via deterministic (ODE/PDE) models, which invariably require laborious parameterisation for each particular process plant.
Physics-Informed Neural Networks/PINN [3-4] combine the best of both worlds: they offer chemistry-compliant NN with proven extrapolation power to revolutionise manufacturing, circumventing parametric estimation uncertainty [5] and enabling efficient process control [6]. Fermentation for ethanol [7] and biopharmaceuticals [8-9] was explored using ML/ANN (not PINN) tools, without embedded first-principles descriptions [3]. Exploring ML/PINN potential for Food+Drink Manufacturing optimisation and control is both possible and timely, hence the topic of this contribution, given tools [6] and datasets from our previous publications available [10-11], with a particular focus on industrial fermentation towards beer production. Correlating feedstock properties with processing conditions and product quality via PINN thus showcases promising technical improvements possible towards process optimisation.
ACKNOWLEDGEMENT
Funding from both the EPSRC (RAPID: ReAl-time Process ModellIng and Diagnostics: Powering Digital Factories, EP/V028618/1) as well as the Royal Society via an International Partnership Grant (2023-25) is gratefully acknowledged.
LITERATURE REFERENCES
[1] Gerogiorgis & Bakalis, Digitalisation of Food+Beverage Manufacturing, Food & Bioproducts Processing, 128: 259-261 (2021).
[2] Lagaris et al., ANNs for solving ODEs and PDEs, IEEE Transactions on Neural Networks, 9(5): 987-1000 (1998).
[3] Karniadakis et al., Physics-informed machine learning, Nature Reviews Physics, 3(6): 422-440 (2021).
[4] Raissi et al., Physics-informed neural networks: A deep learning framework for solving forward+inverse problems involving nonlinear PDEs, Journal of Computational Physics, 378: 686â707 (2019).
[5] Lee et al., Machine learning: Overview of recent progresses and implications for the Process Systems Engineering field, Computers & Chemical Engineering, 114: 111-121 (2018).
[6] Alhajeri, Christofides et al., Physics-informed machine learning modeling for predictive control using noisy data, Chemical Engineering Research & Design,
186: 34-49 (2022).
[7] Pereira et al., Hybrid NN modelling and particle swarm optimization for improved ethanol production from cashew apple juice, Bioprocess & Biosystems Engineering
44: 329-342 (2021).
[8] Peng et al., The ANN approach based on uniform design to optimize the fed-batch fermentation condition: application to production of iturin-A, Microbial Cell Factories, 13: 54 (2014).
[9] Petsagkourakis et al., Reinforcement learning for batch bioprocess optimization. Computers & Chemical Engineering, 133: 106649 (2020).
[10] Rodman & Gerogiorgis, Multi-objective process optimisation of beer fermentation via dynamic simulation, Food & Bioproducts Processing, 100A: 255-274 (2016).
[11] Rodman & Gerogiorgis, Dynamic optimization of beer fermentation: Sensitivity analysis of attainable performance vs. product flavour constraints, Computers & Chemical Engineering, 106: 582-595 (2017).