Definition
For a design problem with two or more objectives f1(x), f2(x), … over
a design-variable vector x, a candidate x* is Pareto-optimal if
there is no other feasible candidate x that beats x* on every
objective simultaneously. The set of all Pareto-optimal candidates is
the Pareto front (or Pareto frontier).
Geometrically, the Pareto front is a curve (two objectives), surface (three), or manifold (more) in objective space. Designs inside the front are dominated — strictly worse than something on the front. Designs on the front represent genuine trade-offs.
Why it beats single-threshold design
Classical engineering design picks a hard threshold on each objective and accepts whatever falls inside. “Sprung-mass acceleration must be under 0.4 g RMS.” “Tyre normal-force variance must be under X.” “The shaking force must be under Y.”
That works when the thresholds are independently informative. It fails when the thresholds interact — a common situation in mechanism design:
- Tightening the transmission angle floor reduces mechanical-advantage risk but pushes the coupler curve away from its target.
- Adding counterweight mass cuts shaking force but raises rotational inertia and packaging volume.
- A heavier crankshaft balances better but costs fuel economy.
A single-threshold design hides the trade-off behind a default. A Pareto-front design surfaces the trade-off as a curve the engineer walks.
In our pipeline
Three places use Pareto fronts directly:
- Geometry × balance co-design for four-bar linkages — link lengths swept against counterweight mass-radius products. Two objectives: coupler-curve precision-point error and shaking-force amplitude. Result: a 2D Pareto front over a 6-variable design space.
- Engine NVH harmonic trade-offs — 1× shaking-force residual against 2× residual under joint counterweight + balance-shaft co-design. The Pareto front shows the cost of reducing 1× to zero (more 2× left over) and vice-versa.
- Quarter-car ride vs handling — sprung-mass acceleration RMS against tyre normal-force variance under spring- rate × damper-coefficient sweep. The classical “softer-or-stiffer” trade-off becomes a sharp curve.
How we compute it
For convex problems, scalarisation (weighted sum, ε-constraint) is sufficient. Our problems are typically non-convex, so we use:
- NSGA-II — non-dominated sorting genetic algorithm, the classical multi-objective evolutionary algorithm. Produces a population of Pareto-optimal candidates per generation.
- Differential evolution with a non-domination ranking — when the underlying single-objective DE workflow is already in place for one of the objectives.
Population sizes are kept small (~50–100) because each candidate requires a full dynamics simulation — the dynamics is the cost bottleneck, not the search.
What the engineer does with it
Pick a point on the front. The picking is the design decision. The optimiser cannot pick — picking encodes the engineer’s understanding of context, market, packaging envelope, regulatory floor, weight budget. The Pareto front gives the engineer a list of candidates that are genuinely on the trade-off and removes the candidates that are not.
See also
- Force balancing — one axis of the geometry × balance Pareto front.
- Engine NVH case study.
- Quarter-car suspension case study.