Robust Design & Design Under Uncertainty

A Structure That Is Optimum Only at One Exact Set of Assumptions May Be a Poor Real-World Design

Deterministic optimisation finds the best design for a single, fixed set of inputs: one load magnitude, one material modulus, one thickness, one set of boundary conditions. The real structure will not see those exact values. Loads vary in magnitude and direction. Material properties scatter around their nominal values. Thicknesses deviate within tolerance. Boundary conditions are never perfectly rigid or perfectly pinned. A design that is sharply optimum at the nominal point may degrade rapidly when any input deviates — and in service, inputs always deviate. Robust design is the practice of seeking designs whose performance varies acceptably across the uncertainty that the real structure will actually experience.

Sources of Uncertainty

Uncertainty enters a structural model from many directions, and a robustness study is only as good as its inventory of what can vary. The common sources are listed below. Note that some are epistemic (we could reduce them with better knowledge) and some are aleatory (irreducible scatter inherent in the process).

• Applied load magnitude — spectra are statistical, not deterministic
• Load direction — real load paths are rarely perfectly aligned with the model
• Boundary conditions — support stiffness is never infinite or zero
• Material properties — modulus, strength, and allowables all scatter
• Thickness and dimensions — within manufacturing tolerance, not at nominal
• Manufacturing tolerances — geometric deviations from the nominal CAD
• Temperature — affecting modulus, strength, and thermal stresses
• Degradation — fatigue, creep, corrosion, and ageing over service life
• Assembly — clamp-up force, preload scatter, joint gap, fastener fit
• Friction and contact — coefficient of friction is poorly controlled
• Preload — bolt tension varies with lubrication, torque method, and fit

Deterministic Design vs Robust Design

A deterministic design approach treats every input as a fixed value and applies a safety factor to the result. The safety factor is meant to absorb the combined effect of all uncertainties, but it is blunt: it does not distinguish a design that is sensitive to a particular variable from one that is not. A robust design approach explicitly models how performance varies with the inputs and seeks a design whose performance distribution is acceptable — not just whose nominal performance has margin. The robust design may have a slightly worse nominal performance but a much narrower spread, which is often the safer engineering choice.

robust-design

Sensitivity: How Fast Does the Response Move?

The foundation of robustness analysis is sensitivity — the rate at which a response of interest changes when an input variable changes. If a small change in a variable causes a large change in the response, the design is sensitive to that variable and will be vulnerable to its uncertainty. Sensitivity is expressed as the partial derivative of the response with respect to the variable. In practice, for non-linear or coupled responses, this derivative is estimated numerically by perturbing each variable in turn and observing the response change.

S_i  =  ∂R / ∂x_i

  where  R   =  response of interest (stress, displacement, frequency, margin)
         x_i =  input variable i  (load, modulus, thickness, …)
         S_i =  sensitivity of R to x_i

  |S_i| large  →  R changes rapidly with x_i  →  design is sensitive to that variable
  |S_i| small  →  R is insensitive to x_i        →  uncertainty in x_i is less concerning

Methods for Exploring Uncertainty

Several methods exist for propagating uncertainty through a structural model, ranging from cheap local checks to expensive global analyses. The right method depends on how non-linear the response is, how many variables matter, and how much computational budget is available.

• Sensitivity analysis — perturb each variable locally; cheap; identifies which variables dominate; misses interactions
• Design of experiments (DoE) — sample the variable space systematically; reveals interactions; moderate cost
• Monte Carlo simulation — draw random samples from each variable distribution and run the model many times; gives a full performance distribution; expensive but general
• Probabilistic analysis — propagates distributions analytically or numerically; efficient for linear problems
• Reliability-based design optimisation (RBDO) — optimises subject to a probability of failure constraint rather than a deterministic margin
• Worst-case analysis — combines all variables at their extremes simultaneously; conservative; can be physically impossible
• Tolerance studies — focus on geometric variation and its effect on fit and load path; essential for assemblies

Mean Performance vs Variability

Robust design is concerned with two properties of the performance distribution, not just one. The mean (or nominal) performance tells you how good the design is on average. The variability — standard deviation, or spread — tells you how much the performance wanders as inputs vary. A design with an excellent mean but large variability is risky: it will frequently produce unacceptable performance. A design with a slightly worse mean but small variability is predictable: it will almost always perform within a known band. Robust optimisation typically minimises an objective that penalises both the mean and the variability, reflecting the engineering preference for predictable performance.

• Mean performance — the expected value of the response across the uncertainty
• Variability — the spread of the response; large spread means unpredictable performance
• A design with good mean and large spread may fail often even though it looks fine on paper
• A design with slightly worse mean and small spread is often the safer engineering choice

Why Design B May Be the Superior Engineering Solution

Consider two candidate designs. Design A is lighter and, at the nominal point, has a higher stiffness. But it sits on a steep part of the response surface: small changes in material modulus or load direction cause large changes in stress. Design B is slightly heavier and slightly less stiff at the nominal point, but its response is flat — the same variations in modulus and load direction barely move the stress. In a deterministic comparison, A wins. In service, where the inputs will never be at their nominal values, B is the design that will reliably meet its requirements. The robustness penalty — the mass or performance you give up to buy insensitivity — is the price of a design that works in the real world rather than on the nominal input sheet.

A robust design carries a penalty — usually a small mass or cost increase — because it is no longer chasing the single sharpest optimum. This penalty is not waste; it is the cost of insensitivity. Present it explicitly in the trade study so the decision to accept or reject robustness is a conscious one, not an accident of which optimisation you ran.

Reliability-Based Design Optimisation

Reliability-based design optimisation (RBDO) formalises the robustness idea by replacing deterministic margins with probabilistic constraints. Instead of requiring stress ≤ allowable / 1.5, RBDO requires that the probability of stress exceeding the allowable be below a specified target — say 10⁻⁴. This ties the margin directly to the uncertainty in the inputs and the consequences of failure, rather than applying a single blanket factor. RBDO is more demanding to set up: it requires distributional information for each variable and a method for estimating failure probability (typically Monte Carlo or a first-order reliability method). It is appropriate where the cost of over-design is high and the data on variability is good.

• Deterministic: stress ≤ allowable / safety factor — blunt, does not reflect actual variability
• RBDO: P(failure) ≤ target probability — ties margin to uncertainty and consequence
• Requires distributional data for each variable — means, standard deviations, distribution type
• Failure probability estimated via Monte Carlo or first-order/second-order reliability methods (FORM/SORM)
• Most valuable when over-design is expensive and variability data is trustworthy

Key takeaways

• A deterministic optimum is the best design for inputs that will never occur exactly; robust design seeks a design whose performance stays acceptable across real-world variation.
• Uncertainty enters from load magnitude and direction, boundary stiffness, material scatter, tolerances, temperature, degradation, assembly and preload — inventory all of them before analysing robustness.
• Sensitivity ∂R/∂xᵢ is the foundation: a large sensitivity means the design is vulnerable to uncertainty in that variable.
• Robust design balances mean performance against variability — a design with a slightly worse mean but small spread is often the safer engineering choice.
• Methods range from cheap local sensitivity analysis to expensive Monte Carlo and RBDO; choose based on non-linearity, number of variables, and budget.
• The robustness penalty — the mass or performance given up to buy insensitivity — is a real cost and should be presented explicitly in the trade study.