2. Automotive simulation landscape

Before we can talk about active learning, we need a clear picture of the simulation disciplines it will be applied to. Automotive CAE is not one monolithic activity; it is a portfolio of physics domains, each with its own governing equations, solver technology, mesh conventions, and — most importantly for this course — its own cost structure. A surrogate model that works beautifully for a steady-state thermal problem may be useless for a transient crash event, because the input-output relationships, the smoothness of the response surface, and the cost per sample are fundamentally different. This chapter maps that landscape.

The Five Core Simulation Disciplines

Crash and impact simulation uses explicit finite element codes (LS-DYNA, Radioss, Abaqus/Explicit) to solve the equations of motion at very small time steps. A full-vehicle frontal impact model typically contains 5–15 million elements, runs for 100–150 ms of physical time, and requires 8–48 hours on 64–256 CPU cores. The output of interest is rarely a single scalar; engineers extract intrusion, acceleration pulses, and injury criteria (HIC, chest deflection) at dozens of locations. The response surface is highly nonlinear and often discontinuous — a 5 mm change in a rail trigger can flip a buckling mode and change peak acceleration by 20 g.

NVH (Noise, Vibration, Harshness) simulation spans modal analysis, frequency response, and acoustic radiation. A body-in-white modal run solves a large eigenvalue problem (10–30 million DOF) and extracts the first 200–500 modes. Frequency response sweeps from 20 Hz to 200 Hz in 1 Hz steps, each step requiring a direct or modal superposition solve. The cost driver here is the number of frequency points and the modal basis size, not the time step. NVH responses are smooth and differentiable, which makes them relatively friendly to Gaussian-process surrogates.

CFD (Computational Fluid Dynamics) for external aerodynamics or underhood cooling solves the Navier-Stokes equations on 50–300 million cells. A steady-state RANS run converges in 4–12 hours on 128–512 cores; a transient DES or LES run can take days. The output is a drag coefficient, lift coefficient, or a surface pressure map. CFD responses are smooth in design space but expensive per sample, and the mesh generation step alone can consume 30–50% of the total turnaround time.

Thermal simulation covers both steady-state and transient heat transfer: battery pack cooling, exhaust system thermal fatigue, and cabin climate. A conjugate heat transfer model couples solid conduction with fluid convection, often requiring co-simulation between a thermal solver and CFD. Steady-state runs are moderate in cost (1–6 hours), but transient thermal cycling over 1000+ seconds of physical time can be as expensive as a crash run.

Durability and fatigue simulation predicts crack initiation and propagation under cyclic loading. It combines a stress analysis (often linear static or modal) with a fatigue post-processor (nCode, FE-SAFE). The stress solve is cheap per load case, but a full durability assessment may require 20–50 load cases, each with a different boundary condition. The output is a fatigue life in cycles, which spans several orders of magnitude and is therefore often analyzed in log space.

Cost Drivers and Fidelity Trade-offs

The cost of a single simulation is driven by four factors: model size (DOF or cell count), physical time or frequency range, solver tolerance, and hardware allocation. These factors interact. Doubling mesh density in a crash model increases cost by roughly 2^4 = 16× because the stable time step shrinks with element size while the element count grows. In CFD, refining the mesh near a separation point improves drag prediction accuracy but can double wall-clock time.

Fidelity trade-offs are therefore not optional; they are the daily reality of CAE. A common pattern is the fidelity ladder: a 1D system-level model for concept phase (seconds per run), a coarse 3D model for design exploration (minutes to hours), and a high-fidelity 3D model for final validation (hours to days). Active learning is most valuable at the middle rung, where each sample is expensive enough to matter but cheap enough to run hundreds of times.

Why the Landscape Matters for Active Learning

Active learning strategies are not interchangeable across these disciplines. Crash simulation benefits from adaptive sampling because the response surface has sharp transitions; a batch of 20 samples can be placed near a suspected buckling threshold. NVH and CFD benefit from Gaussian-process or polynomial-chaos surrogates because their responses are smooth. Thermal transient problems benefit from time-series surrogates that predict the entire temperature curve, not just a scalar. Durability benefits from multi-fidelity methods because the cheap linear-static solve can guide sampling for the expensive full fatigue assessment.

The key insight is that the cost per sample and the smoothness of the response together determine which active learning acquisition function will be effective. A high-cost, rough response (crash) favors expected improvement with a large exploration parameter. A moderate-cost, smooth response (NVH) favors Gaussian-process regression with a small length scale. A very high-cost, smooth response (CFD) favors multi-fidelity or transfer learning approaches.

Practical Implications for Surrogate Modeling

When you build a surrogate for any of these disciplines, the first decision is the design of experiments (DoE). For crash, a Latin hypercube with 50–100 points is a reasonable starting budget; for NVH, 20–30 points often suffice because the response is smooth; for CFD, 15–25 points may be enough if the design space is narrow. The second decision is the surrogate type: Gaussian processes for smooth responses, random forests or gradient-boosted trees for rough or discontinuous responses, and neural networks when the input dimension exceeds 20.

The third decision is the acquisition function. Expected improvement (EI) is the default for single-objective problems. For multi-objective crashworthiness (minimize mass, maximize energy absorption), use expected hypervolume improvement (EHVI). For constrained problems (e.g., intrusion must stay below a threshold), use constrained EI. The choice of acquisition function is where active learning delivers its value: instead of filling the design space uniformly, it places samples where the surrogate is uncertain or where the predicted improvement is high.

What to Watch For

Two failure modes are common when applying active learning to automotive simulation. First, response noise: crash simulations with different mesh densities or solver settings can produce non-physical scatter, which confuses the surrogate. Always verify that the simulation setup is consistent across samples. Second, extrapolation: a surrogate trained on a narrow design space will fail if the optimizer pushes into an unexplored region. Constrain the acquisition function to the training domain, or add a penalty for extrapolation.

With this landscape in mind, the next chapter will introduce the core active learning loop: initial sampling, surrogate training, acquisition, simulation, and update. The cost and smoothness characteristics described here will determine how each step is configured.