A detailed microsimulation of a city district can take hours to run. That is acceptable for a one-off impact study, but it is a poor fit for a planning meeting where someone asks what happens to bus travel times if one junction closes for the afternoon, and then asks again with a different signal plan. Surrogate models address this gap. A surrogate is a machine learning model trained to imitate the outputs of a slow, physics-based simulation, such as average queue length at a junction or expected pressure drop in a water main, so that answers arrive in seconds rather than hours.
The training process starts with the full simulation. Engineers define the inputs that matter, such as demand profiles, signal timings, closures and weather, and then run the detailed model across a designed set of combinations. Each run becomes a labelled example. A regression model or a small neural network learns the mapping from inputs to outputs. The useful result is not a perfect copy of the simulator but a fast estimate with a known error range for the cases it was built to cover.
The main risk is extrapolation. A surrogate trained on weekday traffic will give confident-looking numbers for a holiday event it has never seen, and a planner may not notice. Good practice is to record the input envelope used in training and flag any query that falls outside it. Those queries should be routed back to the full simulation, or at least labelled as indicative only. Teams should also keep a held-out set of detailed runs and check the surrogate against them after every retraining.
Validation does not end at deployment. Live sensor data from the city twin can be compared with surrogate predictions for the same conditions, and persistent gaps signal that road layouts, driver behaviour or demand have changed. When that happens, the model should be retrained and given a new version number, so that a decision made in March can be traced back to the exact model that supported it. Logging each scenario, its inputs, its output and the model version is what turns a clever tool into something an audit committee can accept.
In practice, the most valuable setup is a shared scenario library where operators can compare options side by side, with clear ownership of who approves changes. The surrogate speeds up exploration, but the final call still rests on the detailed model and on human judgement about local conditions. Platforms such as TwinUp's Clavus City are built around this kind of layered workflow, combining fast estimates for exploration with full simulation for decisions that carry real consequences.