Process digital twins for control, state estimation and optimization
A calibrated, physics-based model of your process, kept in step with plant data by state estimation — to estimate what is not measured, test control strategies safely, and feed MPC and real-time optimization.
Book a 30-min technical fit callPhD APC/MPC · CERN · IAV · SkySails
What Is a Process Digital Twin?
A process digital twin is a live, synchronized model of a plant or unit. Unlike a static simulation, it receives real-time data from the DCS or historian and keeps its internal state aligned with the plant through state estimation — so engineers can see what happens inside the process, test changes before making them, and feed model-based control and optimization.
It needs three components: (1) a calibrated dynamic or steady-state process model encoding the physics — heat transfer, reaction kinetics, fluid dynamics; (2) a real-time data connection from the DCS/historian; (3) a state estimator (Kalman filter, moving-horizon estimation or data reconciliation) that keeps the model synchronized with the plant despite measurement noise and model uncertainty.
Three Types of Digital Twins
Asset twin — single equipment item
A pump, compressor, heat exchanger or reactor monitored continuously. The model tracks fouling, degradation and performance, and flags when the asset deviates from the model — before it fails.
Process twin — unit operation or plant section
A distillation column, reactor train or furnace modelled as a thermodynamic system. It supports what-if analysis, optimization of the operating point, testing of control strategies and operator training for start-up and shut-down.
System twin — several units
Several units linked by material and energy balances across a site. It supports site-level scheduling, blending and utility dispatch within physical process constraints.
Problems Solved
No view inside the process
Temperature profiles inside a reactor, concentrations along a column, heat flux at tube walls are not measured directly. A digital twin computes them from first principles and the available measurements, so engineers see the process state, not only the sensor points.
Experiments on the plant are expensive
Trying a new operating point on the real plant risks off-spec product, equipment stress and lost production. On the twin, changes can be tested first, with the process physics respected.
Process knowledge leaves with people
Experienced engineers carry process knowledge that disappears when they leave. A digital twin captures it as a calibrated, documented physics model that the whole team can use.
Model-based controllers drift
APC and RTO degrade as the process changes and models go stale. A twin that continuously reconciles its parameters against plant data keeps the MPC and RTO models accurate.
How We Build It
Process model development
A first-principles thermodynamic, kinetic and flow model built from engineering data — PFDs, datasheets, specifications. Its complexity matches the use case; nothing is over-engineered.
Data connection and reconciliation
Real-time data from the historian (e.g. AVEVA PI, Aspen IP.21, Ignition) via OPC UA or REST. Reconciliation closes mass and energy balances, corrects sensor bias and flags gross errors before they corrupt the model.
State estimation
A Kalman filter, extended Kalman filter or moving-horizon estimator tracks unmeasured states — concentrations, fouling factors, catalyst activity — from the available measurements. At CERN, a nonlinear MHE estimated 5 thermodynamic states of the LHC cryogenic circuit from 3 pressure sensors and temperature at 1 Hz.
Optimization and advisory layer
The live twin feeds an optimizer (RTO, DRTO or economic MPC) that computes optimal operating targets. Operators see the recommendations first — advisory mode, then closed loop only after validation.
Model Development: Getting Complexity Right
When Models Get in the Way
A client once commissioned a dynamic model derived from video-game simulation code. It was elaborate and computationally expensive — but when it was connected to a control loop, the interface matched no standard toolchain, the parameters had no physical meaning, and the controller designed against it performed worse than a hand-tuned PID on the real hardware.
Over-engineered models are a recurring problem in control engineering: they take months to build, break when the physical system changes, are hard to validate against measurements, and often cannot be connected to real control software without a bespoke translation layer.
The answer is not a better complex model. It is starting from the simplest model that captures the behaviour that matters for control.
Tier 1 — Geometry & Basic Physics (Linear)
Derive mass, inertia, spring and damper rates, kinematic constraints and basic aerodynamic coefficients directly from geometry, drawings or CAD. Linearize around the operating point to obtain a state-space model.
- Suitable for Bode analysis, loop shaping, LQR and linear MPC
- Parameters have physical meaning — easy to validate against step responses
- Usually sufficient for feedback design and gain tuning
- Works with any standard control toolchain
Start here. This is enough most of the time.
Tier 2 — Simple Nonlinear Extensions (Simulation)
Add only the nonlinearities the application really needs: actuator saturation, dead band, significant kinematic coupling or dominant nonlinear stiffness.
- Used for closed-loop simulation before hardware deployment
- Parameters still have physical meaning — no black-box fitting
- Light enough for real-time use on standard hardware
- Gives realistic margins the linear model cannot
Add this when controllers from the linear model fail in simulation.
Tier 3 — High-Fidelity Models (Rarely Justified)
Justified only when tiers 1 and 2 cannot capture safety-critical behaviour — for example flexible structures with strong modal coupling, combustion and reaction kinetics, two-phase flow or aeroelastic effects.
- Model complexity becomes part of the research problem
- Needs dedicated identification experiments and validation campaigns
- The longer build time is accepted because simpler tiers are demonstrably insufficient
Use this only with evidence that tier 2 is not enough.
Why Simpler Models Win
Built in days, not months
First-principles derivation is fast. A working linear model of a new system usually takes a few days of engineering — not a six-month modelling project.
Interpretable parameters
Every parameter has a physical meaning — mass, inertia, stiffness, damping — so the model can be validated against measurements and updated when the hardware changes.
Toolchain compatible
State-space models go straight into MATLAB, Python-Control, CasADi and any MPC solver, with no translation layer.
Honest about uncertainty
A simple model known to be approximate is safer than a complex model assumed to be accurate. Knowing the model limits is part of the control design.
Results
See the unmeasured states
Temperature profiles, concentrations and fouling estimated continuously from the measurements you already have.
Test before touching the plant
New operating points, control strategies and start-up sequences tried on the model first, without risking production.
Control models that stay calibrated
Continuous parameter estimation keeps the MPC and RTO models accurate as the plant ages.
Knowledge that stays
Process knowledge captured in a calibrated, documented model instead of in individual heads.
A process digital twin is not a stand-alone product — it connects the other layers: reliable measurements; soft sensors for the gaps (soft-sensor.com); state estimation that keeps the model synchronized; APC/MPC that executes the targets (Industrial Process NMPC / APC); and RTO that computes them (Real-Time Optimization). My work covers the modelling, estimation, control and optimization layers.
Relevant Design Patterns
Book a 30-min technical fit call
If you are evaluating a digital twin project or want to understand which layer of your process to tackle first — a 30-minute call is enough to map the architecture and identify the highest-value entry point.
Book a 30-min technical fit call →