Monte Carlo Method — CAE Glossary

There are three main uses. First is Uncertainty Propagation, which determines the statistical distribution of outputs when there are variations in material properties or load conditions. Second is Reliability Analysis, which evaluates the distribution of failure probability and safety factors. Third is View Factor Calculation for Radiative Heat Transfer, where radiation exchange between complex-shaped surfaces is determined by ray tracing.

Simply put, it's a method that "approximates an integral or expected value that we can't solve directly using the average of random samples." For example, if you want to find the area of a circle, you randomly throw points in a square and estimate the area from the proportion of points that land in the circle—that famous experiment is the essence of the Monte Carlo Method. The name comes from the Monte Carlo Casino in Monaco because the idea of "rolling dice many times" fits perfectly.

With simple Monte Carlo, the convergence rate is $O(1/\sqrt{N})$, so to gain one digit of precision, you need to multiply the sample size by 100. For example, if you want the standard error below 1%, you'd want roughly $N = 10{,}000$ samples.

Consider automotive body structure analysis. Suppose sheet thickness has manufacturing tolerance of ±0.05 mm, Young's modulus varies ±3% between batches, and weld bead position precision is ±1 mm. A deterministic FEM analysis can only input one representative value, but with Monte Carlo, you specify each parameter as a probability distribution (normal or uniform) and run FEM hundreds of times. The result gives you things like the 95th percentile of maximum stress or the coefficient of variation of stiffness.

That's the biggest practical challenge. The standard approach is to create a Response Surface Method (RSM) or kriging surrogate model, then apply Monte Carlo to that. Run the original expensive analysis just tens to hundreds of times to build the surrogate model, then perform 100,000 samplings on the surrogate model.

Uncertainty Propagation wants to know "the full statistical distribution of outputs," while Reliability Analysis focuses on one specific number: "the probability of failure." You define a limit-state function $g(\mathbf{x})$ and calculate the failure probability $P_f$, the probability that $g(\mathbf{x}) \leq 0$.

Good catch. With simple Monte Carlo, estimating $P_f = 10^{-6}$ with 10% coefficient of variation requires about $N \approx 10^8$ samples. That's why practitioners use Importance Sampling, which concentrates samples near the failure region. Alternatively, you can use FORM/SORM approximate methods to find the most critical point (design point), then apply "Subset Simulation" with concentrated Monte Carlo around that point. In aerospace and nuclear sectors requiring $P_f < 10^{-7}$, these techniques are essential.

The 6σ design approach is a typical example. Achieving 6σ quality (3.4 failures per million) requires design that accounts for variability. Use Monte Carlo to find the distribution of FEM results considering tolerances, then confirm that even at the 6σ tail, the design doesn't fail. A classic application is press forming of mass-produced parts: include variability in sheet thickness, friction coefficient, and blank position, then assess the probability of cracks or wrinkles.

Completely different approach. Radiant energy is discretized as numerous rays, and the emission direction, reflection, and absorption of each ray are tracked probabilistically. Launch rays from surface A in random directions and measure what fraction reaches surface B—that fraction directly gives the view factor $F_{A \to B}$.

Spacecraft thermal design is the prime example. In space, convection doesn't exist so radiation dominates, and you need accurate view factors between the complex shapes of solar panels and radiators. Analytical solutions exist only for simple shapes like parallel plates or coaxial discs. For real spacecraft geometries, Monte Carlo is the de facto standard. Thermal analysis software like Thermal Desktop uses this method internally.

Excellent question. With participating media, as a ray travels through the gas, we probabilistically determine if it gets absorbed or scattered based on absorption coefficient $\kappa$. The optical path length $s$ is sampled using Beer's Law $\tau = e^{-\kappa s}$. Compared to Discrete Ordinates Method or P1 approximation, Monte Carlo is more expensive computationally, but handling wavelength dependence and non-grey gas properties is more straightforward—a real strength of the Monte Carlo approach.

Yes, there are Variance Reduction Techniques. Let me introduce the ones most commonly used in practice.

Simple random sampling can accidentally create clumps or voids in the sample distribution. LHS divides each variable's range into N equal strata and guarantees one sample from each stratum—so the input space gets covered uniformly with no gaps. For example, if you can only run FEM 50 times for uncertainty analysis, LHS ensures those 50 samples explore the input space quite evenly. Ansys Workbench's "Design of Experiments" and Altair OptiStruct's robust design modules support LHS as standard.

Sobol Sequence is classified as "Quasi-Monte Carlo." Instead of random numbers, it uses a carefully designed low-discrepancy sequence, so space coverage is extremely efficient. LHS requires you to fix sample size N in advance, but Sobol maintains uniformity even if you add samples incrementally. However, LHS is simpler to implement and has better software support, so I'd recommend starting with LHS, then trying Sobol if accuracy isn't sufficient.

Here's a typical workflow. First, define the probability distributions of input variables (thickness, material constants, loads, etc.). Next, generate 50–200 sample points using LHS and run FEM for each. Build a surrogate model—kriging or Polynomial Chaos Expansion (PCE)—from the input-output data. Finally, perform 100,000 to 1,000,000 Monte Carlo samplings on the surrogate model to get statistical quantities or failure probabilities. Since evaluating the surrogate takes microseconds, 1 million runs finish in seconds.

Surrogate validation is mandatory. Confirm $R^2 > 0.95$ with LOOCV (Leave-One-Out Cross-Validation), or check prediction error on held-out validation points. If the surrogate is wrong, all your Monte Carlo results are meaningless. Recently, Adaptive Sampling and Active Learning techniques automatically place additional samples where the surrogate's uncertainty is largest, improving model fidelity adaptively.