Surrogate models: simulation results in seconds instead of hours
FEM and CFD simulations are among the most compute-intensive steps in vehicle development. We build surrogate models – ML models trained on simulations already computed that approximate results for new design variants in seconds, with stated uncertainty and a clear validity range.
The situation
The simulation department is the bottleneck of early development: every design variant to be evaluated costs a simulation run – and with it hours to days of compute. The design space therefore gets searched more narrowly than would be technically desirable.
Yet the existing simulation archives already hold the knowledge to assess new variants approximately. A surrogate model makes that knowledge usable – as a pre-filter for variant selection, not as a substitute for the verification run.
Typical data basis
The archive of FEM and CFD runs: geometry and load parameters per run plus the associated result quantities or result fields. CAD parameters of the design variants, ideally from a parametric model. Material data and load case definitions.
In addition, a held-out set of runs used exclusively for validation, which never enters training.
Why such initiatives fail
Heterogeneous simulation archives. Different solver versions, meshes and inconsistently documented load cases: preparing a consistent training dataset is the first and often biggest work package.
Fields are harder than scalars. Changing geometries and meshes make predicting entire result fields considerably more demanding than predicting individual scalar quantities.
Spuriously precise numbers. The model has to know its validity range: for variants far outside the training data it must point to a real simulation instead of delivering a spuriously precise number.
Bypassed domain authority. The simulation department keeps technical responsibility. A surrogate model sold as a replacement for the verification run loses exactly the users it needs.
Approach
Scalar quantities. For quantities such as maximum stress, natural frequency or pressure loss we use gradient boosting and Gaussian processes, the latter for their built-in uncertainty estimate. For archives beyond a few thousand runs we use sparse Gaussian process variants. The predicted uncertainty is calibrated against the held-out validation set, so the stated tolerance bands hold up.
Result fields. Predicting stress and deformation fields across changing geometries takes more than an ordinary neural network. With consistent CAD parametrisation, an operator architecture on point clouds is sufficient (DeepONet variants, Fourier Neural Operator); where differing meshes exist, we work geometry-aware on the mesh itself, with graph networks or transformer-based operators such as Transolver. GPU training runs on NVIDIA PhysicsNeMo. Where the customer already uses a commercial CAE AI module (Ansys SimAI, Altair PhysicsAI, Neural Concept), we integrate it instead of building a parallel world – our role is then data preparation, validation and workflow integration.
Crash with a clear boundary. For crash we offer pre-assessment of scalar quantities, such as intrusion values or energy absorption, as a screening step before the regular run. Predicting complete crash fields remains a research topic due to contact, branching and path dependency – it does not belong in a project commitment.
Active learning. Via design of experiments, additional simulation runs are triggered exactly where the model is least certain. The training set grows where it adds the most.
Inside the simulation workflow. The surrogate model sits in the parameter study workflow: variants are pre-assessed in seconds, the promising ones then simulated regularly. If a variant lies outside the validity range, the model outputs no number but a request for simulation.
What you get
Seconds instead of simulation days: design variants can be assessed almost immediately in the early phase. More variants, better decisions: the design space can be searched considerably more broadly before development commits to a few candidates.
Compute where it counts: expensive simulations are reserved for critical variants and final verification runs. Measured by throughput: the value shows in more variants assessed per development loop and higher simulation throughput – not in a blanket time saving.
Typical fields of application: control arms and other chassis components, battery housings, engine mounts, and seat and crash structures.
Where our experience comes from
We predict physical and technical target quantities from existing data – including where new configurations and locations demand reliable generalisation.
Last updated: 13 August 2026
