Polynomial Chaos Expansion (PCE) Simulator

Stochastic Finite Element Methods (SFEM): When material constants (Young's modulus, yield stress, damping), boundary loads or contact stiffnesses are uncertain, engineers want the distribution of responses such as peak stress, maximum displacement or natural frequency. The standard approach is "non-intrusive PCE": reuse the deterministic FEM solver, evaluate it at Gauss-quadrature nodes only, and regress the PCE coefficients. Replacing 10,000 MC runs with around 50 PCE evaluations is routine in automotive, aerospace and nuclear reliability assessments.

CFD and thermofluid uncertainty propagation: Engineers propagate uncertain turbulence-model coefficients (the Cμ of k-ε, for example), inlet velocities and surface roughness through to drag coefficients and temperature distributions. Because each CFD run takes hours or days, MC is essentially impossible — sparse PCE combined with Smolyak sparse grids can characterise the output distribution in 20 to 100 runs. NASA, ONERA and other agencies have many published cases in the aircraft-certification workflow.

Robust design of control and electronics: Sensor noise, resistor tolerances, temperature drift — designers need to know how badly these wobble the control performance or circuit response, without resorting to brutally pessimistic worst-case bounds. Once you have a PCE surrogate, computing Sobol indices on top is cheap, so you can pinpoint which tolerance dominates the variance and feed that into a cost-optimal tolerance allocation. Semiconductor and power-electronics teams have adopted this in recent years.

Climate, geoscience and financial engineering: Sensitivity studies of climate models, uncertain subsurface flow simulations, parameter-uncertain option pricing — large simulations with multi-dimensional random inputs all rely on PCE. In high dimension the tensor-product basis explodes, so practitioners use sparse PCE, adaptive PCE or low-rank tensor decompositions to keep only the basis functions that actually matter.

The biggest trap is using low-order PCE on a non-smooth response. PCE converges exponentially when Y(ξ) is smooth in ξ; but for buckling, impact or contact problems with discontinuities or hard thresholds, the low-order Hermite expansion oscillates like a Gibbs phenomenon. The mean may be right but the variance and tail probability can be badly wrong. The practical test: increase P by 1 and check whether the coefficient norm decays geometrically. If it doesn't, polynomial PCE is the wrong tool and you should switch to sparse grids, hybrid MC, or domain-decomposed PCE.

Next, mismatching the basis to the input distribution. The Wiener-Askey scheme tells you the optimal basis for each distribution (Gaussian → Hermite, uniform → Legendre, exponential → Laguerre, beta → Jacobi). Using Hermite on a uniform input breaks the orthogonality, mixes coefficients and ruins both convergence and accuracy. When several different distributions are involved, transform each input to its standard random variable (Rosenblatt or Nataf transform) before applying the matching basis.

Finally, extrapolating a PCE outside its training range. A PCE is an approximation around the sample cloud (typically ±3σ). Push the input five times further out and Y(ξ) — being a polynomial of ξ — will blow up, but that blow-up is mathematical, not physical. For tail probabilities and extreme-value studies, declare the PCE's region of validity explicitly and supplement it with MC or importance sampling outside that region.