Table of Contents
A surrogate model is a fast statistical approximation of a slow process model, such as a flowsheet simulation, reactor model or CFD study. You run the rigorous model at chosen points, fit a regression or machine-learning model to the results, test it on unseen runs, then use it for optimisation and what-if studies inside its validated range.
Key takeaways
- A surrogate stands in for a simulator, not for the plant. It learns from simulation runs, while a soft sensor learns from plant and lab data.
- It pays off only when the rigorous model is slow and called many times. In an LNG refrigeration case, surrogate-based optimisation found the optimal operating conditions in 2 hours instead of 3 weeks; in another study, a simple ammonia reactor model gained no time reduction.
- About 10 simulator runs per input variable is a reasonable first experiment (Loeppky et al., 2009).
- An optimiser will find the surrogate's errors. Confirm every proposed optimum with the rigorous model and keep the search inside the sampled range.
- Open-source tools cover the full workflow. The hard part is defining the problem, not the software.
What is a surrogate model in chemical process engineering?
Process engineers already own accurate models: a flowsheet in a commercial simulator, a kinetic reactor model, a CFD case of a mixer or burner. These models are slow, and some fail to converge at certain inputs. That is acceptable for one case study, but not when an optimiser, a sensitivity study or a design screening needs thousands of evaluations.
A surrogate, also called a metamodel, replaces the expensive model inside that loop. A review of hybrid modelling in chemical engineering lists this among the ways to combine physics and data: building a faster model of an accurate high-fidelity computer model. Others include correcting a mechanistic model that misses some physics, and inserting a data-driven term for an unknown correlation inside a mechanistic model.
It is easily confused with two related tools:
| Soft sensor | Process surrogate | Formulation model (virtual lab) | |
|---|---|---|---|
| Stands in for | A missing online measurement | A slow simulation | Part of the lab work |
| Trained on | Plant historian and lab results | Runs of a rigorous model | Past recipes and test results |
| Who chooses the data points | The plant's operating history | The engineer, freely | The formulator, at a cost per test |
| Used for | Monitoring and control | Optimisation, design, what-if studies | Choosing the next experiments |
| Main risk | Drift and sensor faults | Inherits simulator errors; extrapolation | Extrapolation to new chemistry |
Our guides to soft sensors and to the virtual lab for industrial R&D cover the other two columns.
When does a surrogate model pay off?
The benefit comes from replacing many slow calls.
Savage and colleagues applied dimensionality reduction and an array of Gaussian processes, modelling both the objective and the feasibility constraints, to the CryoMan cascade LNG refrigeration cycle. Their approach located the set of optimal operating conditions in only 2 h as opposed to the original 3 weeks.
Granacher and colleagues replaced three process unit simulations inside a superstructure optimisation: a biomass-to-Fischer-Tropsch fuel process, an ammonia reactor and a co-electrolysis unit. The most complex model was approximated in less than 10% of the original simulation time, co-electrolysis needed 67%, and for the simple ammonia reactor model there was no time reduction. The authors expect savings of up to 50% for simulations that take more than about 10 seconds per run.
The rule is simple: if one run of your model takes milliseconds, optimise it directly. A surrogate earns its cost when runs take seconds to hours, when the model is called thousands of times, or when convergence failures stall the optimiser.
Which type of surrogate model should you choose?
Neural networks are especially common in chemical process modelling, while polynomial and kriging-based approaches are also applied in process design. The table summarises the main options.
| Model type | Strengths | Limits | Example tools |
|---|---|---|---|
| Polynomial (response surface) | Transparent, easy to embed in an optimiser | Misses strong nonlinearity | Any statistics package; PySMO |
| Sparse algebraic regression | Short equations that engineers can read and optimisers can handle | Needs a licence | ALAMO |
| Gaussian process (kriging) | Returns a prediction and its uncertainty | Loses efficiency beyond a few dozen inputs | scikit-learn, PySMO |
| Neural network | Many outputs, strong nonlinearity, large data sets | Needs more runs; harder to inspect | Keras with OMLT |
| Gradient-boosted trees | Robust, little tuning | Piecewise, so no smooth gradients | OMLT |
A few points from the documentation behind these tools:
- The IDAES process modelling platform can integrate ALAMO, PySMO or Keras surrogate models into IDAES flowsheets. It describes ALAMO as a regression tool for equation-based surrogates that requires a separate licence.
- scikit-learn's Gaussian processes give probabilistic predictions with confidence intervals, but lose efficiency in high-dimensional spaces, when the number of features exceeds a few dozen.
- OMLT is an open-source package that inserts neural network and gradient-boosted tree surrogates into larger optimisation problems written in Pyomo. Schweidtmann and Mitsos showed that networks embedded this way can be solved by deterministic global optimisation, with test cases that include a chemical process and a compressor plant.
How do you build a surrogate model, step by step?
- Define the decision. Write down what the surrogate is for, for example "maximise hydrogen yield subject to a nitrogen limit". The decision fixes which outputs matter and how accurate they must be.
- Choose few inputs and honest ranges. Every extra input raises the number of runs. Use process knowledge or a quick sensitivity run on the rigorous model to drop inputs with little effect. Set ranges to what the plant can actually reach.
- Generate a space-filling design. Latin hypercube sampling splits each input range into equal intervals and places exactly one point in each. Start with about 10 runs per input.
- Run the simulator in batch and log everything. Record the inputs, every output, the convergence status and the run time. Runs that fail to converge show where the feasible region ends; do not delete them silently.
- Fit a baseline, then a more flexible model. Keep the flexible one only if it beats the baseline on runs it has not seen.
- Validate against the decision. Check the error of each output on held-out runs, and check physical consistency: mass balance closure, mole fractions that sum to one, no negative flows.
- Add runs where the model is weakest. Use the model's uncertainty to choose the next batch of simulations, then refit.
- Optimise, then confirm. Run the rigorous model at the proposed optimum. If the two disagree, add that point to the data and repeat.
The IDAES examples show this workflow on an autothermal reformer that produces syngas for a fuel cell system. The surrogate has two inputs, the natural gas bypass fraction and the natural gas to steam ratio, and predicts steam flow, reformer duty and outlet composition. Of 2,800 points simulated from a rigorous flowsheet, the notebook trains on a random 100 with an 80/20 split, embeds the Keras network in the flowsheet and maximises hydrogen production with nitrogen kept below 34 mol%.
How many simulation runs do you need?
Loeppky, Sacks and Welch studied the informal rule that an initial computer experiment needs about 10 runs per input dimension, using Gaussian process models and Latin hypercube designs. They concluded that n = 10d is a reasonable rule of thumb for an initial experiment for well-behaved problems. When 10 runs per input give low accuracy, the rate of improvement with more runs also tends to be small, which can call for reformulating the problem.
In practice, that means 50 to 100 runs for a 5 to 10 input problem as a first pass, and a decision point: if accuracy is poor, narrow the ranges, split the operating region or remove inputs before launching thousands of runs.
Adaptive sampling helps after the first batch. Granacher and colleagues started from 10 simulated points and added batches of 10 where the model was most uncertain. Their active learning gave better predictions than random sampling for the same database size, with test mean squared errors below 2% in all three cases.
How do you optimise with a surrogate without being misled?
An optimiser searches for the best value the surrogate predicts. If the surrogate is wrong somewhere, for example near the edge of the sampled range, the optimiser will often go exactly there. Three safeguards are common.
Fall back to the simulator when uncertainty is high. In the Granacher study, a surrogate prediction was used only when its predicted uncertainty was below a threshold of 8%; otherwise the original simulation ran and its result was added to the training data.
Take small, checked steps. Pyomo's trust region solver handles problems that mix equation-based models with black-box models. It replaces the expensive models with low-fidelity functions and moves in small, incremental steps towards an optimal solution, accepting a step only after checking it, which means fewer calls to the expensive functions.
Model the constraints, not only the objective. In the LNG study, separate Gaussian processes modelled the feasibility constraints alongside the objective. A surrogate that predicts profit but not whether the compressor stays within its limits will propose infeasible operating points.
Surrogate-based optimisation is a family of methods, also called model-based derivative-free optimisation. A recent tutorial chapter compares Bayesian optimisation, linear and quadratic approximation methods, tree-based and radial basis function methods on reactor case studies. For experiments rather than simulations, see our comparison of design of experiments and Bayesian optimisation.
Pure data fit or hybrid model?
A pure surrogate knows nothing about conservation laws. A hybrid model keeps the mechanistic structure, such as mass and energy balances, and learns only the uncertain part, such as a reaction rate or an efficiency. Hybrid modelling, physics-informed machine learning and model calibration are the three families the hybrid modelling review compares on a chemical reactor case study.
Commercial simulators have moved in this direction. In May 2021, AspenTech announced first-principles-driven hybrid models with AI embedded in Aspen HYSYS and Aspen Plus, and reduced-order hybrid models for its planning and dynamic optimisation products. Product names and packaging change, so check what your current licences include.
Choose a pure surrogate when you need speed inside a loop and the rigorous model is trusted. Choose a hybrid when the mechanistic model itself does not match plant data, for example because kinetics or fouling are poorly known. For a model that tracks one operating asset, see digital twins in manufacturing.
What we have seen at KABAS AI
Our Virtual Lab is built to test formulations and process parameters virtually before the lab or pilot phase. The two deployments we describe publicly are formulation projects rather than flowsheet surrogates, but the discipline is the same.
At ADFC (Sonatrach Group), comparisons between simulated and actual results guide refinements, and validated formulations are stored for reuse. At ENAD Shymeca, historical formulation data was migrated into the platform before it was used for prediction, and the case study reports a 40% reduction in physical testing cycles.
Three practical lessons carry over to process surrogates:
- Start from the decision, not the model. Our AI formulation optimisation guide shows how to turn a specification into numbers first.
- Confirm before acting. Every prediction that drives a decision should be checked against the reference, whether that is the lab, the pilot plant or the rigorous simulator.
- Keep the models where the know-how is. A process model and its surrogate encode how your plant runs. Virtual Lab can be deployed on-premise or on a local PC with no internet connection, and our guide to sovereign, on-premise AI covers the options.
Our chemistry solutions page shows how this fits R&D and process optimisation in chemical companies.
Sources
- Savage et al., An integrated dimensionality reduction and surrogate optimization approach for plant-wide chemical process operation, AIChE Journal (Cambridge repository)
- Granacher, Kantor and Maréchal, Increasing Superstructure Optimization Capacity Through Self-Learning Surrogate Models, Frontiers in Chemical Engineering, 2021
- Loeppky, Sacks and Welch, Choosing the Sample Size of a Computer Experiment: A Practical Guide, Technometrics 51(4), 2009
- Bradley et al., Perspectives on the Integration between First-Principles and Data-Driven Modeling, LAPSE preprint, 2021
- IDAES documentation: Surrogate modeling tools
- IDAES examples: Keras flowsheet optimization (autothermal reformer)
- Ceccon et al., OMLT: Optimization and Machine Learning Toolkit, Journal of Machine Learning Research 23, 2022
- Schweidtmann and Mitsos, Deterministic Global Optimization with Artificial Neural Networks Embedded, arXiv 1801.07114
- scikit-learn documentation: Gaussian Processes
- SciPy documentation: scipy.stats.qmc.LatinHypercube
- Pyomo documentation: Trust Region Framework Method Solver
- Neufang et al., Surrogate-Based Optimization Techniques for Process Systems Engineering, arXiv 2412.13948, 2024
- AspenTech Expands Application of Industrial AI to Achieve New Profitability and Sustainability Goals, May 2021 (via CIMdata)
Frequently asked questions
What is the difference between a surrogate model and a reduced-order model?
The terms overlap. Surrogate model and metamodel usually mean a data-fitted approximation built from runs of a rigorous simulator. Reduced-order model often means a simplified model derived from the physical equations, although vendors also use the term for data-fitted models. What matters in practice is what the model was trained on and over which range it was validated.
How many simulation runs do I need to build a surrogate model?
A common starting point is about 10 runs per input variable, so 50 runs for five inputs. Published work on computer experiments supports this as a first experiment. If accuracy is still poor at that size, adding runs usually helps slowly, and it is better to reduce the number of inputs or narrow the ranges.
Can a surrogate model replace the process simulator?
No. The surrogate is trained on the simulator's results, so it inherits the simulator's assumptions and errors and only knows the range that was sampled. Keep the rigorous model to confirm every optimum the surrogate proposes, and rebuild the surrogate when the flowsheet, thermodynamic model or equipment changes.
Is a surrogate model the same as a soft sensor?
No. A soft sensor is trained on plant historian data and lab results to estimate a variable that is not measured online. A process surrogate is trained on runs of a simulation model to stand in for that model during optimisation or design studies. Both are data-driven models, but their data, purpose and failure modes differ.
Which software can we use to build process surrogates?
Open-source options cover the whole workflow: SciPy for sampling designs, scikit-learn for Gaussian processes and other regressors, and the IDAES platform with Pyomo and OMLT for embedding surrogates in flowsheet optimisation. Commercial process simulators also offer hybrid and reduced-order modelling features, so check what your existing licences include.
Can surrogate models run on our own servers?
Yes. Surrogate models are small and need no internet connection once trained. Because they encode process know-how, many companies keep both the simulation data and the models on their own infrastructure.



