1. Course overview

High-fidelity CAE simulation sits at the center of modern product development. A single nonlinear structural solve, a transient CFD run, or a coupled thermal–structural analysis can take anywhere from minutes to many hours, and design optimization routinely demands hundreds or thousands of such evaluations. When a genetic algorithm explores a 12-variable design space with a population of 40 over 60 generations, that is 2,400 solver calls — a workload that can consume weeks of compute time and stall a design cycle. Surrogate models (also called metamodels or response surfaces) address this bottleneck by learning an inexpensive approximation of the input–output relationship from a limited set of expensive simulations, then substituting that approximation inside the optimization loop.

This course builds the practical competence to construct, validate, and deploy such surrogates within a CAE-driven design workflow. The first chapter establishes the conceptual map: where surrogates sit in the design loop, why they are necessary, and which engineering challenges determine whether they succeed or fail.

The Design Loop and Where the Surrogate Enters

A conventional simulation-driven optimization loop has four stages: parameterize the geometry, run the CAE solver, evaluate the objective and constraints, and update the design variables. The surrogate does not replace the solver; it inserts a fast, learned layer between design variables and responses, so that the optimizer can query thousands of candidate designs at negligible cost. The expensive solver is then reserved for generating the training data and for final verification of the optimized design.

The figure below shows this architecture as a signal-and-data flow: design variables x feed both the high-fidelity solver (for sampling) and the trained surrogate (for optimization queries), with a verification path that returns the surrogate-optimal design to the solver before acceptance.

Design Variables x = (x₁ … xₙ) bounds, parameterization Design of Experiments LHS / Sobol / CCD N samples in design space e.g. N = 120 High-Fidelity CAE FEA / CFD / thermal cost: minutes–hours per evaluation Training Set {(xᵢ, yᵢ)}, i = 1…N y = stress, drag, temp… Surrogate Model Kriging / RBF / GPR cost: milliseconds Optimizer GA / NSGA-II / SQP thousands of queries Verification re-run CAE on optimum check error < tolerance if error too large: refine / resample sample evaluate fit predict

Why Surrogates Are Necessary

The motivation is fundamentally economic. Consider a vehicle crashworthiness study where one explicit dynamic simulation takes 4 hours on 32 cores. A 10-variable optimization with a population-based algorithm needs roughly 1,500 evaluations — about 6,000 core-hours, or 250 wall-clock days on a single 32-core node. With a surrogate trained on 150 solver runs (600 core-hours), the optimization itself completes in minutes, and only the final candidate requires one additional verification run. The trade is explicit: a bounded, quantifiable approximation error in exchange for a two-to-three order-of-magnitude reduction in computational cost.

Three further motivations reinforce this. First, exploration breadth: surrogates let engineers screen a design space coarsely before committing to detailed analysis. Second, noise and non-smoothness: solver outputs can carry mesh-dependent noise or discrete contact events that confuse gradient-based optimizers; a smooth surrogate regularizes the response. Third, multi-disciplinary coupling: when structural, thermal, and aerodynamic solvers must be coordinated, a unified surrogate layer can evaluate all responses consistently and cheaply.

Key Challenges That Determine Success

The apparent simplicity of "fit a model to data" conceals several engineering difficulties that this course addresses directly.

  • Curse of dimensionality. Sample requirements grow rapidly with the number of design variables. A 5-variable problem may be well served by 80–150 Latin Hypercube samples; a 30-variable problem may need thousands, eroding the cost advantage. Dimensionality reduction, variable screening, and adaptive sampling are essential countermeasures.
  • Accuracy versus cost. The surrogate must be accurate enough that the optimizer's optimum is not an artifact of approximation error. Validation metrics — R², RMSE, and especially leave-one-out cross-validation error — must be computed and interpreted, not merely reported.
  • Extrapolation risk. A surrogate is only trustworthy inside its training domain. Optimizers are notorious for pushing variables to the edges of, or beyond, the sampled region. Constraining the search to the convex hull of samples, or using uncertainty-aware models such as Gaussian process regression, mitigates this.
  • Constraint handling. Feasibility boundaries (yield stress, maximum temperature) are often where accuracy matters most, yet data there may be sparse. Infill criteria that target the constraint boundary are a recurring theme.
  • Data quality and consistency. Mesh convergence, solver tolerances, and geometry parameterization must be fixed across all samples; otherwise the surrogate learns solver noise rather than physics.
A useful rule of thumb when scoping a surrogate study: budget at least 10 samples per design variable for a first-pass Kriging model, then use cross-validation error to decide whether to add samples or reduce dimensionality. If the 10×N budget already exceeds the cost of direct optimization, the surrogate approach is not justified for that problem.

A Concrete First Example

To make these ideas tangible, consider a cantilever bracket with three design variables: web thickness t ∈ [2, 8] mm, fillet radius r ∈ [3, 15] mm, and hole diameter d ∈ [6, 20] mm. The objective is to minimize mass subject to maximum von Mises stress ≤ 180 MPa. A finite element solve takes about 90 seconds.

A practical workflow proceeds as follows. Generate 60 samples using a Latin Hypercube design over the three-variable box. Run the 60 FE analyses (about 90 minutes total) and record mass and peak stress for each. Fit a Gaussian process regression model with a Matérn 5/2 kernel, standardizing inputs to zero mean and unit variance. Evaluate with 10-fold cross-validation: if the stress model achieves R² ≥ 0.95 and RMSE ≤ 8 MPa, proceed; otherwise add 20 samples in the region of largest cross-validation error. Then run NSGA-II against the surrogate for 5,000 evaluations, extract the Pareto front, and re-run the FE solver on the three most promising candidates. Accept a candidate only if its solver-evaluated stress is within 5% of the surrogate prediction and satisfies the constraint.

Typical failure symptoms at this stage are instructive. If cross-validation R² is high but the verified optimum violates the stress constraint, the surrogate is likely extrapolating near the feasibility boundary — the remedy is constraint-focused infill sampling. If predictions scatter randomly against solver results, check that mesh density and solver settings were identical across all 60 runs. If the model is accurate but the optimizer returns designs clustered at variable bounds, the sampling plan may not have covered the corners of the design space adequately.

What This Course Delivers

Across 30 chapters, the course moves from this conceptual foundation through sampling strategies, surrogate families (polynomial response surfaces, radial basis functions, Kriging/Gaussian processes, support vector regression, and neural network surrogates), validation methodology, adaptive infill, multi-fidelity and multi-objective formulations, and finally deployment within commercial CAE environments. Each technique is presented with the inputs it requires, the computations it performs, and the observable criteria by which its output should be judged. The goal is not familiarity with terminology but the ability to build a surrogate that an engineering team can trust enough to make design decisions with.

← Previous 📖 Contents Next →