13.4 Optimization methods in vibration design

Optimization for Vibration Design

Optimization methods in vibration design let engineers systematically find the best configuration of a mechanical system rather than relying on trial and error. The goal is to tune design parameters so that dynamic performance (natural frequencies, mode shapes, frequency response) meets targets while staying within real-world constraints like stress limits and manufacturing tolerances.

These problems require three ingredients: an objective function to minimize or maximize, a set of design variables the engineer can change, and constraints the solution must satisfy. The choice of optimization algorithm then depends on the nature of the problem, whether it's smooth or discontinuous, single-objective or multi-objective, and how expensive each function evaluation is.

Performance Criteria and Design Variables

The objective function captures what you're trying to achieve. Common objectives include:

• Minimizing total mass for a given stiffness target
• Maximizing the lowest natural frequency to push resonance out of the operating range
• Minimizing peak frequency response amplitude at a critical location

Design variables are the quantities the optimizer is free to change. These typically include:

• Structural dimensions (cross-section thickness, flange width, rib spacing)
• Material properties (density, elastic modulus when selecting among candidate materials)
• Damping coefficients (dashpot values, viscoelastic layer thickness)

Constraints keep the solution physically realistic:

• Allowable stress levels (e.g., von Mises stress $\leq$ 250 MPa)
• Displacement limits (e.g., maximum deflection $\leq$ 5 mm under operating loads)
• Frequency exclusion zones to avoid resonance (e.g., no natural frequency between 1000–2000 Hz for a rotating machine running at those speeds)

Multi-objective Optimization and Sensitivity Analysis

Most real vibration design problems involve conflicting goals. Minimizing weight tends to reduce stiffness, which lowers natural frequencies. Multi-objective optimization finds the Pareto front, the set of solutions where you can't improve one objective without worsening another. The engineer then picks the best trade-off from that set.

Sensitivity analysis tells you how much each design variable influences the response. For example, increasing a beam's thickness from 10 mm to 12 mm (a 20% change) might raise the first natural frequency by roughly 15%. Variables with high sensitivity deserve the most attention during optimization; low-sensitivity variables can sometimes be fixed at nominal values to reduce problem size.

Robustness matters too. A design that's optimal at exactly 25°C but fails at 80°C isn't useful. Robust designs account for:

• Operating condition variability (temperature swings, load fluctuations)
• Manufacturing tolerances (e.g., $\pm$ 0.1 mm on machined dimensions)

Formulating Vibration Design Problems

Problem Definition and Objective Function

Good formulation is half the battle. A poorly posed optimization problem will give you a mathematically "optimal" answer that makes no engineering sense.

The standard form looks like this:

1. Define the objective. Express it mathematically in terms of design variables. Example: minimize total mass $m_{total}(x)$ subject to the first natural frequency $\omega_1(x) \geq 100$ Hz.
2. Write the constraints as equality or inequality expressions. Inequality constraints are more common in vibration design (e.g., $\sigma_{max} \leq \sigma_{allow}$, $|x_{max}| \leq 5$ mm).
3. Connect objectives and constraints to vibration theory. The relationship between design variables and dynamic response often comes from eigenvalue problems. For a simple undamped system, the natural frequency satisfies $\omega_n^2 = \frac{k}{m}$, but for complex structures you'll rely on FEA-based eigenvalue solves. The Rayleigh quotient provides an upper-bound estimate of the fundamental frequency and is useful for approximate or early-stage optimization.

Design Variables and Parameterization

Each design variable needs a feasible range grounded in practical limits. A beam thickness variable might be bounded to 5–20 mm because thinner stock isn't available and thicker stock won't fit the assembly.

Parameterization reduces the number of design variables when the geometry is complex. Instead of treating every node's position as independent, you might describe a curved surface with a polynomial having just a few coefficients. Fewer variables means faster convergence and a smoother design space.

Your choice of algorithm also shapes how you set up the problem:

• Gradient-based methods need smooth, differentiable functions. If your objective has discontinuities (e.g., a discrete material choice), these methods struggle.
• Heuristic algorithms (genetic algorithms, particle swarm) handle discrete variables and non-smooth landscapes but require many more function evaluations.

Uncertainty and Robust Optimization

Deterministic optimization assumes you know every parameter exactly. In practice, you don't. Material properties like Young's modulus can vary by $\pm$ 5% from lot to lot. Operating temperatures might range from 20°C to 80°C.

Robust optimization accounts for this by optimizing not just the nominal performance but also the worst-case or statistical spread. A common approach is to minimize the mean response while constraining the standard deviation, so the design performs well across the full range of uncertain inputs. Reliability-based optimization goes further, ensuring the probability of constraint violation stays below a target (e.g., 99.9% probability that stress stays below the allowable limit).

Optimization Algorithms for Vibration Design

Gradient-based and Heuristic Methods

Gradient-based methods use derivative information to move efficiently toward the optimum. They converge quickly on smooth problems but can get trapped in local minima.

• Sequential Quadratic Programming (SQP): Solves a quadratic subproblem at each iteration to handle nonlinear constraints. A workhorse for smooth, constrained vibration problems.
• Newton's method and quasi-Newton variants (e.g., BFGS): Use second-order (or approximate second-order) information for fast local convergence.

Heuristic (metaheuristic) methods explore the design space more broadly and don't require gradient information.

• Genetic algorithms (GA): Evolve a population of candidate designs through selection, crossover, and mutation. Good for discrete or mixed-variable problems.
• Particle swarm optimization (PSO): Candidate solutions "swarm" through the design space, sharing information about promising regions.

How to choose: if the problem is smooth, convex, and has a single basin of attraction, gradient-based methods are faster. If the landscape has multiple local optima, discrete variables, or is highly nonlinear, heuristic methods are more reliable (though slower).

Constrained and Multi-objective Optimization

Most vibration design problems have constraints, so you need techniques that enforce them:

• Lagrange multipliers convert a constrained problem into an unconstrained one by introducing multiplier variables for each constraint. The multiplier values at the optimum tell you the "cost" of tightening that constraint.
• Penalty methods add a term to the objective function that grows large when constraints are violated, steering the optimizer away from infeasible regions.

For multi-objective problems, specialized algorithms generate the Pareto front:

• NSGA-II (Non-dominated Sorting Genetic Algorithm II) is the most widely used. It ranks solutions by dominance and maintains diversity along the Pareto front.
• MOEA/D decomposes the multi-objective problem into a set of single-objective subproblems solved simultaneously.

Computational Efficiency and Sensitivity Analysis

Each function evaluation in a vibration optimization may require a full FEA eigenvalue solve, which can be expensive. Surrogate models replace the expensive simulation with a cheap approximation trained on a limited set of runs:

• Response surface methodology (RSM): Fits a polynomial (often quadratic) to sampled data. Simple and fast, but less accurate for highly nonlinear responses.
• Kriging (Gaussian process regression): Interpolates exactly through sampled points and provides an uncertainty estimate, which helps decide where to sample next.

For gradient-based optimization, you need derivatives of the objective and constraints with respect to design variables. Three common approaches:

1. Finite difference approximation: Perturb each variable by a small amount and re-solve. Easy to implement but requires $n+1$ (forward difference) or $2n$ (central difference) extra solves for $n$ variables.
2. Adjoint variable method: Computes gradients with respect to all design variables in roughly the cost of one extra solve, regardless of $n$. Very efficient for problems with many design variables.
3. Analytical sensitivity: Derived directly from the governing equations. Most accurate but requires significant derivation effort for each new problem type.

Interpreting Optimization Results

Analysis of Optimal Solutions

Once the optimizer converges, the work isn't done. You need to verify that the result makes engineering sense.

• Check physical feasibility. An optimal beam thickness of 8.5 mm is useful only if your manufacturing process can produce it. Round to the nearest available stock size and re-evaluate performance.
• Assess sensitivity near the optimum. A perturbation analysis might show that a 1% change in material density shifts the natural frequency by only 0.5%, meaning the design is relatively insensitive to that parameter. That's a good sign for robustness.
• Benchmark against the baseline. Quantify the improvement: "The optimized design achieves a 20% mass reduction while maintaining the same first natural frequency as the original."

Constraint Analysis and Practical Considerations

Look at which constraints are active (at their limit) versus inactive (satisfied with margin). Active constraints drive the design. If the stress constraint is active but the displacement constraint has plenty of margin, further improvement requires either relaxing the stress limit or changing the problem formulation.

Post-optimization verification should include:

1. Modal analysis of the final design to confirm natural frequencies and mode shapes match predictions.
2. Frequency response analysis across the operating range to check for unexpected resonances the optimization might have missed (especially if only a few modes were tracked).
3. Transient or harmonic analysis under realistic loading to validate stress and displacement predictions.

Practical implementation factors that the optimizer doesn't "see" also matter:

• Minimum feature sizes for the chosen manufacturing process (e.g., 1 mm for CNC milling)
• Material cost versus performance gain
• Maintenance access and inspection requirements

Communication and Decision Making

Optimization results are only valuable if they inform real decisions. When presenting to stakeholders:

• Show the trade-offs explicitly. A table or Pareto plot showing that a 5% increase in material cost buys a 15% improvement in vibration isolation makes the decision concrete.
• Compare the top candidate designs side by side with quantitative metrics (mass, natural frequencies, peak stress, cost).
• Justify the recommended design by connecting the optimization outcome to project requirements and long-term benefits like reduced fatigue risk or extended service life.