Food Science & Technology · Postgraduate & Research · also Mathematical Modelling in Food Processing, Computer-Aided Food Process Engineering, Process Optimisation
Building and using models of food processes: kinetic models of quality and microbial change, numerical solution of heat and mass transfer, computational fluid dynamics, multi-objective optimisation, data-driven and machine-learning models, and digital twins of food plants.
In the exams: PhD research; ICAR AIEEA-PG Code 18 Unit II (system analysis, automation and online data acquisition)
A model is a simplified description of a system that can be used to explain or predict its behaviour. Food engineers have always used models, from the first-order equations of microbial death behind every sterilisation process to the psychrometric charts used to design dryers. What has changed is scale and speed: computers now solve coupled equations of heat, mass and momentum transfer in realistic geometries, fit kinetic models to large data sets, search thousands of process settings for the best compromise between safety, quality and cost, learn patterns from plant data, and run alongside real plants as digital twins. A food technologist who understands how models are built, what they assume and how far they can be trusted can use them to design processes faster and more safely; one who does not risks believing whatever a colourful simulation shows.
This subject covers the types of model (empirical, mechanistic and data-driven); kinetic modelling of quality, nutrient and microbial changes; numerical methods, chiefly finite differences and finite elements; modelling heat transfer in thermal processing; modelling simultaneous heat and mass transfer in drying and frying; computational fluid dynamics; process simulation software; single- and multi-objective optimisation; machine learning and artificial intelligence in food processing; uncertainty and sensitivity analysis; and digital twins and Industry 4.0 in food plants. It builds on mathematics-for-food-technologists, food-transport-phenomena and the unit-operation subjects, and connects with statistics-experimental-design-food for parameter estimation and response surfaces.
Process modelling is the working method of much doctoral research in food engineering and is examined in postgraduate entrance examinations in food process engineering, which include system analysis, automation and online data acquisition, and in research-fellowship examinations. The practice questions test understanding of model types, assumptions and validity, with short calculations of dimensionless numbers, kinetic and secondary models, numerical stability limits, drying models, sensitivity coefficients and optimisation.
Models are classified in several ways. Empirical (black-box) models describe the relationship between inputs and outputs observed in experiments without representing the mechanisms, for example a polynomial fitted to describe how the moisture of a dried product depends on air temperature and time, or the response-surface models of designed experiments; they are quick to build and accurate within the range of data, but they should not be extrapolated. Mechanistic (white-box, first-principles) models are derived from physical, chemical and biological laws, such as conservation of mass, energy and momentum, Fourier's and Fick's laws, and reaction kinetics; their parameters have physical meaning (thermal conductivity, diffusivity, activation energy), they can be extrapolated with more confidence, and they explain behaviour, but they need more knowledge and effort and still rest on simplifying assumptions. Data-driven models, including machine learning, learn relationships from large data sets; they can capture complex non-linear behaviour without explicit mechanisms but need much data and can fail badly outside their training range. Hybrid or grey-box models combine a mechanistic structure with empirical or data-driven parts, for example a heat-transfer model in which an unknown heat-transfer coefficient is estimated from data, and are common in practice.
Other distinctions matter. Deterministic models give one output for given inputs, while stochastic models include randomness, for example variability in spore heat resistance or initial contamination. Steady-state models describe conditions that do not change with time, such as a continuous evaporator at constant operation, while dynamic models describe change with time, such as heating of a can or start-up of a dryer. Lumped-parameter models assume uniform properties within a body, for example a well-mixed tank or a small particle with negligible internal resistance, and are described by ordinary differential equations; distributed-parameter models allow properties to vary in space and are described by partial differential equations.
Model development follows a cycle: define the purpose and the required accuracy; identify the system, its boundaries and the important phenomena; state assumptions explicitly (for example constant properties, one-dimensional transfer, no shrinkage); formulate the equations with initial and boundary conditions; estimate parameters from literature or experiments; solve the equations analytically or numerically; verify and validate; and apply the model, revising it when new data reveal its limits. Verification asks whether the equations are solved correctly, for example by checking a numerical solution against an analytical solution or by refining the mesh until results stop changing; validation asks whether the right equations were chosen, by comparing predictions with independent experimental data not used to estimate the parameters. A model can be perfectly verified and still invalid.
Dimensionless numbers help decide which phenomena matter and simplify models. The Biot number, Bi = hL/k, compares external convective to internal conductive resistance: for Bi below about 0.1, internal temperature gradients are negligible and a lumped model is adequate; for large Bi, the surface temperature quickly reaches the medium temperature. The Fourier number, Fo = αt/L², is dimensionless time for conduction. The Reynolds number decides whether flow is laminar or turbulent, and the Prandtl, Nusselt and Sherwood numbers appear in heat- and mass-transfer correlations. Model quality is judged by goodness of fit (the root mean square error and the coefficient of determination R², though R² alone can mislead for non-linear models), by the randomness of residuals, by the precision and correlation of parameter estimates, and, when comparing models with different numbers of parameters, by criteria such as the Akaike information criterion that reward fit but penalise complexity. The principle of parsimony favours the simplest model adequate for the purpose.
Kinetic models describe how a quality attribute or a microbial population changes with time under given conditions. For many chemical changes, the rate is modelled as dC/dt = −kCⁿ, with zero-order (n = 0, constant rate, as for some browning and oxidation indices), first-order (n = 1, exponential decay, as for many vitamins and pigments) and occasionally second-order kinetics. Microbial inactivation is classically modelled as first order, log₁₀(N/N₀) = −t/D, with the decimal reduction time D; but many survivor curves show shoulders or tails, which the Weibull model, log₁₀(N/N₀) = −(t/δ)ᵖ, describes with a scale parameter δ and a shape parameter p (p > 1 shoulder, p < 1 tail). Biphasic models describe subpopulations of different resistance.
Microbial growth is modelled by primary growth models that describe the lag phase, exponential growth and stationary phase. The Baranyi and Roberts model, widely used in predictive microbiology, represents the lag by the physiological state of the cells and gives three main parameters: the lag time, the maximum specific growth rate μ_max and the maximum population density; the modified Gompertz model is an older empirical alternative. Growth is usually modelled on log counts. These primary models are fitted to data at constant conditions, and their parameters are then related to the environment by secondary models.
Secondary models describe how primary-model parameters depend on temperature, pH, water activity and other factors. For inactivation, the D value is related to temperature by the z value, log₁₀(D/D_ref) = (T_ref − T)/z, or the rate constant by the Arrhenius equation, k = k_ref exp[−(Ea/R)(1/T − 1/T_ref)]; writing it with a reference temperature within the experimental range, rather than with the pre-exponential factor, reduces the strong correlation between estimated parameters and improves estimation. For growth, the Ratkowsky square-root model relates the growth rate to temperature as √μ_max = b(T − T_min) below the optimum, where T_min is a theoretical minimum growth temperature; for example, with b = 0.03 (h^-0.5 per °C) and T_min = −2 °C, the predicted μ_max at 8 °C is (0.03 × 10)² = 0.09 per hour. Cardinal parameter models use minimum, optimum and maximum values of each factor, and the gamma concept multiplies separate terms for temperature, pH and water activity. Tertiary models package these in software and databases that predict growth or inactivation for user-specified conditions.
Parameters are estimated by non-linear regression, minimising the sum of squared residuals. In the two-step approach, primary models are fitted to each isothermal experiment and the resulting parameters are then fitted by the secondary model; in the one-step (global) approach, all data are fitted simultaneously with the combined model, which uses the data more efficiently and gives more realistic confidence intervals. Confidence intervals and correlation matrices of parameters should always be reported. Real processes are not isothermal: temperature changes during heating, cooling and storage. Predictions for dynamic conditions integrate the rate equation over the temperature history, for example ln(C/C₀) = −∫k(T(t)) dt for first-order kinetics, which is the basis of process lethality and cook-value calculations and of shelf-life prediction from time-temperature records. Non-isothermal experiments, in which samples are heated with a known temperature profile, can also be used to estimate kinetic parameters, provided the design gives enough information. Kinetic models for quality and safety are combined to optimise processes (chapter 8): because microbial inactivation usually has a much higher activation energy than quality loss, high-temperature short-time conditions favour quality.
Most models of real food processes cannot be solved analytically, because of irregular shapes, temperature-dependent properties, phase changes and coupled phenomena, so they are solved numerically. In the finite difference method, the domain is divided into a grid of nodes and derivatives are replaced by differences between nodal values derived from truncated Taylor series. For one-dimensional transient conduction, ∂T/∂t = α∂²T/∂x², the explicit (forward-time, centred-space) scheme gives the new temperature at node i as T_i' = T_i + Fo_m(T_(i+1) − 2T_i + T_(i−1)), where Fo_m = αΔt/Δx² is the mesh Fourier number. The scheme is simple but conditionally stable: for interior nodes Fo_m must not exceed 0.5, otherwise errors grow and the solution oscillates; convective boundaries impose a stricter limit. For example, with α = 1.4 × 10⁻⁷ m²/s, typical of many moist foods, and Δx = 1 mm, the time step must be at most Δx²/(2α) = 3.6 s. The implicit (backward Euler) scheme evaluates the spatial derivative at the new time level, giving a system of equations to solve at each step; it is unconditionally stable, allowing larger time steps, but is first-order accurate in time. The Crank-Nicolson scheme averages the two and is second-order accurate in time, though it can show oscillations with large steps.
Boundary conditions are applied at the edges: prescribed temperature (Dirichlet), prescribed flux or symmetry (Neumann, zero gradient at a centre line), and convection to a medium (Robin, −k∂T/∂x = h(T_s − T_∞)). Accuracy depends on grid spacing and time step; a grid-convergence study, refining the mesh until the result changes negligibly, is part of verification. Finite differences suit simple geometries such as slabs, cylinders and spheres, and they are the basis of many heat-penetration and drying models.
The finite element method divides the domain into small elements of simple shape (triangles or quadrilaterals in two dimensions, tetrahedra or hexahedra in three), within which the unknown is approximated by simple functions (shape functions) of the values at the element's nodes. The governing equation is expressed in an integral (weak) form, and assembling the contributions of all elements gives a system of algebraic equations. Finite elements handle irregular geometries, such as a chicken leg, a pear or a loaf, and variable properties naturally, with meshes refined where gradients are steep, and they are widely used in heat transfer, mass transfer and the mechanics of food materials. Commercial and open-source multiphysics packages provide finite element solvers, mesh generators and post-processing.
The finite volume method divides the domain into control volumes and applies conservation laws to each, ensuring that mass, energy and momentum are conserved exactly over each volume; it is the basis of most computational fluid dynamics codes (chapter 6). Whatever the method, good practice includes a mesh independence study, checking against analytical solutions for simple cases, reporting the time step and convergence criteria, using realistic property data (often temperature- and composition-dependent, estimated from composition by correlations such as those of Choi and Okos), and validating against experiments. Common errors include too coarse a mesh near surfaces, unstable explicit time steps, wrong units of property data, and boundary conditions that do not represent the real process, such as assuming an infinite heat-transfer coefficient when the surface heat transfer actually limits heating.
Thermal processes for canned and packaged foods must deliver a target lethality at the slowest-heating point while limiting quality loss, and models predict the temperature at that point and its lethal effect. For solid and highly viscous foods heated by conduction, the transient heat equation is solved for the container geometry, usually a finite cylinder for cans, a slab for trays and pouches, or irregular shapes for whole products, with a convective boundary condition representing heat transfer from the retort medium. The slowest-heating point is the geometric centre. Analytical series solutions exist for simple shapes with constant properties; numerical solutions allow temperature-dependent properties and variable retort temperatures. The semi-empirical formula method developed by Ball characterises heat-penetration curves by the heating rate index f_h (the time for the straight-line portion of the semi-log plot to traverse one log cycle) and the lag factor j_h, parameters that are the first term of the series solution and allow process times for other retort temperatures and initial temperatures to be calculated; numerical models now complement it.
For liquids and liquids with particles, heat is transferred by natural convection in still containers or forced convection in agitated ones. In natural convection, hot liquid rises along the wall and cooler liquid sinks in the centre, so the slowest-heating zone moves towards the bottom of the container, about 10 to 15 per cent of the height from the base in many cases; CFD models (chapter 6) have shown how this zone changes with time and viscosity. Rotating retorts (end-over-end or axial rotation) induce forced convection and shorten processes. In continuous aseptic processing of particulate foods, models must predict the temperature at the centre of the fastest-moving, slowest-heating particle, which depends on the residence time distribution and the fluid-to-particle heat-transfer coefficient, both uncertain, which is why such processes are validated with conservative assumptions and biological indicators.
The temperature history predicted at the cold point is converted into lethality, F = ∫10^((T − T_ref)/z) dt, using the z value of the target organism, and into a cook value, C = ∫10^((T − T_ref)/z_q) dt, using a z value for quality (often 20 to 40 °C or more, compared with about 10 °C for spores). Because quality attributes are degraded throughout the container, not only at the cold point, the volume-average quality retention is computed by integrating the kinetic model over the temperature field. Models then compare processes: a higher retort temperature or a thinner pack shortens the process and improves average quality for the same cold-point lethality, which is why pouches and trays can give better quality than cans. Variable retort temperature profiles, which change the medium temperature during the process, are an optimisation problem (chapter 8).
Newer heating methods require coupled models. In ohmic heating, an electric current passes through the food, generating heat internally at a rate equal to σE², where σ is the electrical conductivity and E the electric field strength; because σ rises with temperature and differs between liquid and particles, models must couple the electric field and heat transfer, and particles less conductive than the carrier fluid may heat more slowly, a key safety issue. Microwave and radio-frequency heating generate heat by dielectric losses; the power absorbed per unit volume depends on the frequency, the dielectric loss factor and the local electric field, which is non-uniform because of standing waves and edge and corner effects, so microwave models solve Maxwell's equations coupled with heat transfer, and simple models use Lambert's law of exponential decay of power with depth, characterised by the penetration depth. Such models explain cold spots and runaway heating and guide the design of packages, foods and ovens. All thermal process models used for safety decisions must be validated with temperature measurements and, where appropriate, microbiological challenge tests, and regulators and process authorities expect conservative assumptions.
Drying removes water by simultaneous heat transfer to the product and mass transfer of water from it. Drying curves show a constant-rate period, in which the surface is wet and evaporation is controlled by external conditions (air temperature, humidity and velocity), the surface staying near the wet-bulb temperature; and one or more falling-rate periods, in which internal moisture movement, by liquid diffusion, capillary flow and vapour diffusion, limits drying and the product temperature rises. Many foods show little or no constant-rate period. Falling-rate drying is commonly modelled with Fick's second law, ∂X/∂t = D_eff∇²X, using an effective diffusivity D_eff that lumps together all internal transport mechanisms and depends on temperature (often by an Arrhenius relation) and moisture. For a slab of half-thickness L dried from both faces with constant D_eff and negligible external resistance, the moisture ratio MR = (X − X_e)/(X₀ − X_e) approaches, for longer times, the first term of the series solution, MR ≈ (8/π²) exp(−π²D_eff t/(4L²)), so D_eff is estimated from the slope of ln MR against time. Typical effective diffusivities for fruits and vegetables lie roughly between 10⁻¹¹ and 10⁻⁹ m²/s.
Thin-layer drying models are simpler empirical or semi-theoretical equations fitted to drying curves of single layers: the Lewis (Newton) model MR = exp(−kt), the Henderson and Pabis model MR = a exp(−kt), and the Page model MR = exp(−ktⁿ), among many others; for example, with the Page model and k = 0.01 min⁻ⁿ, n = 1.2, the moisture ratio after 60 minutes is exp(−0.01 × 60^1.2) = 0.30. They are useful for describing and comparing drying but do not extrapolate well to other conditions or geometries. More complete models couple heat and mass transfer (as in the Luikov equations), include shrinkage, which changes the diffusion path and is often significant for fruits and vegetables, account for the moisture dependence of properties, and model the product as a porous medium with liquid, vapour and air phases. Such models predict not only moisture but also temperature, and therefore quality changes, case hardening, and the energy efficiency of dryers.
Deep-fat frying is a process of intense simultaneous heat and mass transfer. The food is immersed in oil at about 160 to 190 °C; water at the surface evaporates violently, producing bubbles that increase the external heat-transfer coefficient; as frying proceeds, a dry crust forms at the surface, while the core remains moist and below about 100 °C, cooking by conduction. An evaporation front moves inward, separating the crust, where temperature rises above 100 °C and Maillard browning and acrylamide formation occur, from the core. Models of frying therefore often treat it as a moving-boundary problem, with heat conduction in the crust and core, water evaporation at the front, and vapour transport through the crust; some models use a porous-media approach for liquid water, vapour and oil. The heat-transfer coefficient between oil and food changes during frying with the intensity of bubbling and is a key, often fitted, parameter.
Oil uptake is an important quality and health issue, and models and experiments show that most oil enters the product during cooling after removal from the fryer, when the vapour inside the pores condenses, creating a pressure drop that draws surface oil into the crust; uptake therefore depends on the crust's microstructure, the amount of surface oil, and the draining and cooling conditions, and less on frying time. Models predict moisture loss, crust thickness, core temperature and, coupled with kinetics, colour and acrylamide formation, and are used to choose oil temperatures, frying times and pre-treatments (such as blanching, drying or coatings) that balance texture, colour, oil content and safety. Vacuum frying, at reduced pressure and lower temperature, and air frying, which uses hot air with little oil, have been modelled in similar ways. As with drying, validation against measured moisture and temperature profiles is essential because parameters such as diffusivities, heat-transfer coefficients and crust properties are uncertain.
Computational fluid dynamics (CFD) solves the equations of fluid flow, the continuity equation (conservation of mass), the Navier-Stokes equations (conservation of momentum) and usually the energy equation, together with equations for species transport, on a computational mesh, most often by the finite volume method. It predicts velocity, pressure, temperature and concentration fields in equipment where measurements are difficult. Laminar flows, common in viscous food processing, can be solved directly. Turbulent flows, common in air handling, cleaning, mixing and many heat exchangers, are usually solved with Reynolds-averaged Navier-Stokes (RANS) models, which compute time-averaged flow and represent turbulence with additional equations; the k-ε model is robust for many flows, and the k-ω SST model handles near-wall flows and separation better. Large eddy simulation resolves larger turbulent eddies directly and models only the smallest, at much higher computational cost. Non-Newtonian fluids, such as purees, doughs and sauces, require rheological models (power law, Herschel-Bulkley) for viscosity.
Additional models extend CFD to food problems. Multiphase models include volume-of-fluid methods, which track the interface between immiscible fluids (for example liquid filling or free surfaces in tanks), Euler-Euler methods for dispersed phases, and Euler-Lagrange (discrete phase) methods that track individual droplets or particles through the flow, used for spray dryers to predict droplet trajectories, drying histories and wall deposition. Porous-media models represent packed beds, stacked produce in cold stores and grain bulks. Radiation models are needed for ovens and infrared heating. Setting up a CFD model involves choosing the geometry and simplifications, generating a mesh (with fine layers near walls, where the dimensionless wall distance y+ must suit the chosen turbulence treatment), specifying boundary conditions (inlet velocities and temperatures, outlets, wall conditions), material properties, models and solver settings, and checking convergence of residuals and of key quantities.
CFD is applied throughout food processing. In thermal processing, it has visualised natural convection in cans and pouches, the movement of the slowest-heating zone, and flow and temperature distribution in retorts and pasteurisers, including the holding tubes of continuous processes, where residence time distributions affect lethality. In refrigeration, it predicts airflow and temperature in cold stores, refrigerated display cabinets, refrigerated containers and pallet loads of fruit, guiding the design of vented packaging and the placement of products so that all are cooled. In drying, it models air distribution in tray and tunnel dryers and the behaviour of droplets in spray dryers. In ovens, it predicts air velocity and heat transfer, explaining uneven baking. In mixing and fermentation, it predicts flow patterns, shear rates, gas dispersion and mixing times in stirred tanks and bioreactors, supporting scale-up. In cleaning, it predicts wall shear stress, which relates to the removal of soils and biofilms in CIP, and identifies dead zones in fittings and equipment where cleaning may fail, supporting hygienic design.
The attractiveness of CFD images can hide errors, so validation is essential. Predictions are compared with measurements such as velocity fields by particle image velocimetry or hot-wire anemometry, temperature histories from thermocouples or wireless loggers, residence-time distributions from tracer experiments, and cleaning results. Sensitivity of results to mesh, turbulence model, boundary conditions and property data is checked, and uncertainty in inputs (for example the heat load of produce or the porosity of a pallet) is assessed. CFD models are also used to build faster reduced-order models and to generate data for machine-learning surrogates, which can run quickly enough for optimisation or for use in digital twins (chapter 11). Computational cost remains a constraint for large, transient, multiphase problems, so the level of detail is matched to the question.
Process simulators calculate mass and energy balances, equipment sizes, utility requirements and sometimes costs for whole processes, represented as flowsheets of unit operations connected by streams. Steady-state flowsheet simulators were developed for the chemical and petroleum industries and are used in food processing for continuous operations such as evaporation, distillation of ethanol, oil refining, sugar processing and drying. Each unit operation (mixer, heat exchanger, evaporator, separator, dryer, reactor) is represented by a model, and the simulator solves the system, either unit by unit in sequence with iterative handling of recycle streams (sequential modular approach) or as one large set of equations (equation-oriented approach). Accurate simulation needs property models: for water and steam, standard steam tables are built in, but food materials are complex mixtures whose properties (specific heat, density, thermal conductivity, boiling-point elevation, viscosity) are estimated by composition-based correlations, such as those of Choi and Okos, which compute properties from the mass fractions of water, protein, fat, carbohydrate, fibre and ash and temperature, or entered from measurements. Food components are often defined as pseudo-components for this purpose.
Many food processes are batch or semi-continuous, such as fermentation, cheese making, brewing and many ingredient manufacturing processes. Batch process simulators model sequences of operations in vessels with time, scheduling of equipment, cleaning and turnaround times, resource use (labour, utilities, materials) and throughput, identifying bottlenecks and estimating capital and operating costs, which is valuable in designing new plants and in techno-economic analysis of new products such as precision-fermented proteins or upcycled ingredients.
Detailed models of individual operations are built in multiphysics software, which couples heat transfer, mass transfer, fluid flow, electromagnetics and structural mechanics, usually with finite element methods, and in CFD packages (chapter 6), both commercial and open-source. General-purpose programming and numerical environments, used through scripting languages, are widely used by researchers to write custom models, solve differential equations, fit parameters, run optimisations and Monte Carlo simulations, and build machine-learning models, with the advantage of full control and reproducibility of code. Predictive microbiology tools and databases provide growth and inactivation models for pathogens and spoilage organisms under specified conditions, and specialised tools exist for thermal process calculation, refrigeration design, packaging permeability and life-cycle assessment.
Simulation results are only as good as their inputs and models. Good practice includes: defining the purpose and boundaries of the simulation; checking mass and energy balances for closure; using validated property data and stating their sources; calibrating key parameters, such as heat-transfer coefficients or evaporator fouling factors, against plant data; testing the sensitivity of results to uncertain inputs; validating predictions with measurements; and documenting assumptions, versions and settings so that others can reproduce the work. Simulation becomes especially powerful when linked with economics (capital and operating costs, net present value) for techno-economic analysis, and with environmental assessment (energy, water and emissions) for life-cycle assessment, so that process alternatives can be compared on cost, quality and sustainability before money is spent on equipment. Integration with plant data systems, so that models are updated with real operating data, leads to the digital twins discussed in chapter 11.
Optimisation finds the values of decision variables that give the best value of an objective function while satisfying constraints. In food processing the decision variables might be temperatures, times, flow rates, formulation proportions or equipment sizes; objectives might be cost, energy use, throughput, nutrient retention or a quality score; and constraints might include a minimum lethality, a maximum moisture content, legal composition limits, equipment capacities and physical bounds. A problem is linear if the objective and constraints are linear functions of the variables, and non-linear otherwise; unconstrained or constrained; continuous or with integer (discrete) variables; and static or dynamic, when the decisions are functions of time.
Linear programming solves linear problems efficiently, for example by the simplex method, and is the classic tool for least-cost formulation of feeds, sausages, flour blends and beverages, in which ingredient proportions are chosen to meet nutrient and quality specifications at minimum cost. For instance, to blend two flours costing 20 and 30 currency units per kilogram with 10 and 14 per cent protein to obtain a blend with at least 12 per cent protein at minimum cost, the optimum is the blend that just meets the protein constraint, half of each, at 25 currency units per kilogram. Non-linear problems are solved by gradient-based methods, such as steepest descent, Newton and quasi-Newton methods, and constrained problems by methods based on Lagrange multipliers and the Karush-Kuhn-Tucker conditions, such as sequential quadratic programming. Gradient-based methods converge to a local optimum near the starting point, which may not be the global optimum if the problem has several. Response surface methodology optimises empirically, fitting second-order models to designed experiments and finding the stationary point or using desirability functions to combine several responses into one score between 0 and 1. Derivative-free and global methods include evolutionary algorithms such as genetic algorithms, which evolve a population of candidate solutions by selection, crossover and mutation; particle swarm optimisation; and simulated annealing, which accepts occasional worse solutions to escape local optima. These are robust for complex, noisy or discontinuous problems but need many evaluations of the model.
Dynamic optimisation chooses time-varying controls. In thermal processing, variable retort temperature profiles, in which the retort temperature changes during the process, have been optimised to minimise process time or maximise surface or average quality retention while delivering the required lethality; studies found that for conduction-heated foods the gain in average quality over the best constant temperature is often modest, while gains in surface quality or time can be larger. In drying, air temperature and humidity profiles are optimised to reduce energy and quality loss; in fermentation, feeding profiles in fed-batch culture are optimised to maximise productivity. Dynamic optimisation methods include control vector parameterisation, in which the time profile is represented by a few parameters, and approaches based on Pontryagin's maximum principle.
Most food process decisions involve trade-offs between conflicting objectives: higher sterilisation temperatures improve safety margins and throughput but may reduce surface quality; more drying reduces water activity but uses energy and darkens colour; cheaper ingredients reduce cost but may lower sensory quality. Multi-objective optimisation does not give a single answer but the set of Pareto-optimal solutions, those for which no objective can be improved without worsening another; plotted, they form the Pareto front, from which decision makers choose according to their priorities. Methods to generate the front include the weighted-sum method (optimising a weighted combination of objectives, varying the weights, which cannot find non-convex parts of the front), the ε-constraint method (optimising one objective while constraining the others at successive levels), and multi-objective evolutionary algorithms that find many Pareto points in one run. Optimisation results depend entirely on the model; if the model is wrong or used outside its validated range, the 'optimum' may be unsafe or poor, so optimised conditions must be confirmed experimentally and, for safety-related processes, with margins for uncertainty (chapter 10).
Machine learning (ML) builds models that learn patterns from data rather than from explicit equations. Supervised learning relates inputs to known outputs: regression predicts continuous values (such as product moisture from process conditions, or shelf life from composition) and classification assigns categories (such as defective or acceptable, authentic or adulterated). Unsupervised learning finds structure without labels, through clustering of similar batches or dimensionality reduction. Reinforcement learning learns actions that maximise a reward through interaction, and is studied for process control. Common algorithms include linear and regularised regression; decision trees and ensembles such as random forests and gradient-boosted trees, which perform well on tabular process data; support vector machines; k-nearest neighbours; and artificial neural networks, in which layers of connected nodes with non-linear activation functions are trained by adjusting weights through backpropagation to minimise prediction error. Deep learning uses many layers and specialised architectures: convolutional networks for images and spectra, recurrent and attention-based networks for time series and sequences.
A sound workflow begins with the question and the data. Data are collected, cleaned (handling missing values, outliers, sensor errors and time alignment), and engineered into useful features; they are split into training, validation and independent test sets, the last representing future use (for example later production campaigns, not random rows from the same days). Models are trained on the training set, their settings (hyperparameters) tuned by cross-validation, and their final performance reported on the test set, using appropriate metrics (root mean square error and R² for regression; accuracy, sensitivity, specificity and the confusion matrix for classification). Overfitting, in which a flexible model memorises noise, is controlled by more data, simpler models, regularisation and early stopping. Data leakage, when information from the test data or from the future seeps into training, gives misleadingly good results and is a common error.
Applications of ML in food processing are growing quickly. Soft sensors predict quantities that are hard or slow to measure, such as moisture, viscosity or fermentation biomass, from routinely measured temperatures, pressures, flows and spectra, allowing real-time control. Image-based systems grade produce, detect defects and foreign bodies, and monitor baking or roasting colour (see advanced-analytical-omics). Predictive maintenance uses vibration, current and temperature data to predict equipment failure. Quality prediction relates raw-material properties and process settings to product quality, supporting adjustment of settings for each batch. Demand forecasting, scheduling and supply-chain optimisation use ML on commercial data. In research, ML models accelerate formulation, for example predicting texture or flavour from composition, and act as fast surrogates of expensive mechanistic or CFD models for optimisation.
Hybrid models combine the strengths of mechanistic and data-driven approaches: a mechanistic model provides structure that respects physical laws, and ML corrects its errors or estimates uncertain parameters from data; physics-informed neural networks include the governing equations in the training objective. Hybrid models generally extrapolate better and need less data than pure ML. Limitations of ML must be respected. Models can only be as representative as their training data; changes in raw materials, equipment or recipes cause data drift, so models must be monitored and retrained. Many ML models are hard to interpret, which matters when decisions affect safety or must be explained to regulators; explainability methods, such as feature-importance measures and Shapley-value-based explanations, help but do not make a model causal. Correlation learned from plant data is not causation, so ML suggestions for changing settings should be tested. For safety-critical functions, such as determining a thermal process, ML may support but should not replace validated mechanistic models and established safety margins.
Every model prediction is uncertain. Sources include parameter uncertainty (kinetic parameters, thermal properties and heat-transfer coefficients estimated with error), natural variability (differences between batches of raw material, initial contamination, container fill and headspace, spore heat resistance), model-structure uncertainty (simplifying assumptions, choice of kinetic model), input uncertainty (process conditions that fluctuate) and measurement error in the data used for fitting and validation. Distinguishing variability, which is real heterogeneity that cannot be reduced by more measurement, from uncertainty, which reflects lack of knowledge and can be reduced, matters in risk assessment.
Uncertainty is propagated through models to the outputs by several methods. Analytical (first-order) propagation uses the partial derivatives of the output with respect to each input, as in the propagation of measurement uncertainty: for independent inputs, the variance of the output is approximately the sum of the squared sensitivities times the input variances. It is quick but approximate for non-linear models and large uncertainties. Monte Carlo simulation samples each uncertain input from its probability distribution (normal, log-normal, uniform, triangular or empirical), runs the model for each combination, typically thousands of times, and builds the distribution of the outputs, from which means, standard deviations and percentiles (such as the 5th and 95th) are obtained. Correlations between inputs must be represented, and the number of iterations must be enough for stable results, especially for the tails of the distribution, which matter most for safety. For example, a Monte Carlo analysis of a thermal process may show that, given variability in heat penetration and spore resistance, the lethality delivered to the slowest-heating containers falls below target in a small fraction of cases even when the nominal calculation meets it, which informs safety margins. The same approach underlies quantitative microbial risk assessment (see food-risk-analysis).
Sensitivity analysis asks which inputs most influence the output, so that effort can be directed to measuring them better, controlling them more tightly, or simplifying the model by fixing unimportant ones. Local sensitivity analysis evaluates the partial derivative of the output with respect to each parameter at a nominal point, often as a normalised sensitivity coefficient S = (∂y/y)/(∂p/p), the percentage change in output per percentage change in the parameter. For example, if a 10 per cent increase in thermal diffusivity changes the predicted process time by −8 per cent, the normalised sensitivity is about −0.8. One-at-a-time variation of each parameter by a set amount is a simple version. Local methods are informative near the nominal point but can miss interactions and non-linear effects over the full range.
Global sensitivity analysis explores the whole range of inputs simultaneously. Screening methods such as the Morris method identify which of many inputs are important at modest computational cost, by computing elementary effects along random trajectories through the input space. Variance-based methods, such as Sobol indices, decompose the variance of the output into contributions from each input and from interactions: the first-order index gives the share of output variance due to an input alone, and the total-effect index includes all its interactions. Regression-based measures, such as standardised regression coefficients or rank correlation coefficients from Monte Carlo samples, are simpler indicators. Results often show that a few inputs dominate; in thermal processing models, for example, the heat-transfer coefficient at the surface matters little once it is large, while the thermal diffusivity and the z value can matter greatly. Reporting uncertainty and sensitivity alongside predictions is good practice and is increasingly expected in regulatory submissions and publications. Decisions can then be made with explicit margins, or, where needed, more data can be collected on the inputs that matter.
Industry 4.0 refers to the integration of digital technologies into manufacturing: networks of sensors and connected devices (the industrial internet of things), cyber-physical systems in which physical processes and computation are linked, cloud and edge computing, big-data analytics, artificial intelligence, advanced robotics and automation. In food plants it builds on long-established layers of automation, often described by the ISA-95 model: sensors and actuators at the field level; programmable logic controllers and distributed control systems that run processes; supervisory control and data acquisition (SCADA) systems and human-machine interfaces for operators; manufacturing execution systems (MES) that manage production orders, batch records, quality data, traceability and performance (such as OEE); and enterprise resource planning (ERP) systems for business functions. Industry 4.0 links these layers and the supply chain through data, using open communication standards (such as OPC UA) so that equipment from different suppliers can share information.
Typical benefits in food processing include real-time monitoring of critical control points and quality parameters with automatic records, process analytical technology such as in-line spectroscopy for continuous quality control, soft sensors and advanced process control, predictive maintenance, energy and water monitoring, flexible production of smaller batches, and end-to-end traceability, sometimes using distributed ledgers. Challenges include the cost of sensors and integration in older plants, hygienic design of sensors in wet and washdown environments, data quality and standardisation, cybersecurity of connected control systems, the skills needed to manage data and models, and ensuring that automated decisions remain validated and compliant with food-safety requirements.
A digital twin is a virtual representation of a physical asset, process or product that is connected to it by data, updated as the real system changes, and used to monitor, predict and optimise its behaviour. It is often distinguished from a digital model, which is not connected to live data, and a digital shadow, which receives data from the physical system but does not feed decisions back. A digital twin of a process has several components: sensors and data streams from the plant; models, mechanistic, data-driven or hybrid, that represent the process and estimate unmeasured states; methods to assimilate data and update model parameters (for example state estimation with Kalman filters); and interfaces that present predictions and recommended actions to operators or send them to control systems.
Examples in the food sector are emerging. A digital twin of a retort or pasteuriser can compute the delivered lethality for each batch from measured temperatures and models of heat penetration, detect deviations and support release decisions. A dryer twin can estimate product moisture in real time and adjust air conditions to save energy while meeting specifications. A fermentation twin can estimate biomass and product concentration from off-gas and other data and optimise feeding. In the cold chain, digital twins of fruit consignments use temperature records and mechanistic models of fruit cooling, respiration and quality decay to estimate the remaining shelf life of each pallet, supporting decisions about routing and selling, and more broadly a product twin can follow a batch through the supply chain. Building a useful twin requires a clear purpose, reliable data, models validated over the operating range and maintained as the plant changes, and organisational processes that act on its outputs. The value comes not from the technology itself but from better decisions: fewer deviations, less waste, lower energy use and faster, safer innovation. The modelling methods in the earlier chapters of this subject are the foundation on which such systems are built.
A research student uses this subject to build a model for a thesis: choosing between empirical and mechanistic approaches, writing the governing equations with boundary conditions, solving them by finite differences or a multiphysics package, estimating parameters by non-linear regression with confidence intervals, validating against experiments, and analysing sensitivity. Examination questions test model types, Biot and Fourier numbers, stability limits, kinetic secondary models, drying models and optimisation concepts.
Five years into a process engineering or R&D role, a technologist uses simulation to size equipment and utilities, CFD to diagnose uneven cooling in a cold store or oven, kinetic models to design shelf-life tests and thermal processes, optimisation to reduce cost and energy, and ML soft sensors from plant data, always checking predictions against measurements.
A senior engineering or innovation leader decides on investments in simulation capability, automation and digital twins, sets standards for model validation in safety-critical decisions, uses techno-economic models to screen new processes, and manages the data, skills and cybersecurity needed for Industry 4.0.
A researcher at this level develops new models of coupled transport and transformation in foods, hybrid physics-ML methods, efficient optimisation under uncertainty, and digital-twin frameworks for food plants and supply chains, and contributes to validation standards that let models support regulatory decisions.
This manual introduces process modelling, simulation and optimisation for food engineers at postgraduate level. It does not teach the use of particular software packages, derive numerical methods rigorously, or give validated models for specific products and equipment; parameter values and thresholds quoted are typical and must be determined for each application. Statistical parameter estimation is treated in statistics-experimental-design-food, transport theory in food-transport-phenomena and thermal process calculation in thermal-processing-canning. Industry 4.0 and digital twins are developing fields, and the examples describe common approaches rather than mature standards.
Original practice questions written to the syllabus and pattern of the exams named in each question's examTags. They are not past papers and do not reproduce any board's questions.
These notes are free to read here, with no account.
What the Laisha app adds:
Free for ten days, then a subscription. We would rather write that here than have you find it later.