Design of Experiments and Optimization - Theory & Concepts

Design of Experiments and Optimization

Introduction to Design of Experiments (DoE)

Design of Experiments (DoE) is a systematic, rigorous approach to engineering problem-solving that applies principles and techniques at the data collection stage to ensure valid, defensible, and supportable engineering conclusions. Instead of changing one variable at a time (OFAT), DoE involves changing multiple variables simultaneously to understand their individual and interactive effects on a process or product.

Efficiency: DoE extracts the maximum amount of information from the minimum number of experimental runs.
Interaction Effects: Unlike the traditional One-Factor-At-A-Time (OFAT) approach, DoE can identify interactions where the effect of one variable depends on the level of another. For example, the effect of curing temperature on concrete strength might be drastically different depending on the water-cement ratio used.
Optimization: DoE helps identify the optimal settings for multiple variables to achieve the best possible outcome (e.g., maximum strength, minimum cost).

Interact with the simulation below to explore the Design of Experiments methodology.

Response Surface Methodology (RSM) Optimization

Adjust the Temperature and Pressure variables to find the optimal settings that maximize the Material Strength.

Temperature (°C)20

Pressure (atm)1.0

Predicted Material Strength

0.0 MPa

Full Factorial Design

A full factorial design consists of two or more factors (independent variables), each with discrete possible values or "levels," and whose experimental units take on all possible combinations of these levels across all such factors.

Identify Factors: Determine the independent variables you want to test (e.g., Temperature, Pressure).
Determine Levels: Decide on the high ( $+$ ) and low ( $-$ ) settings for each factor (e.g., Temperature at $20^\circ C$ and $40^\circ C$ ).
Create the Design Matrix: For a $2^k$ design (where $2$ is the number of levels and $k$ is the number of factors), list all possible combinations. A $2^2$ design has $4$ runs; a $2^3$ design has $8$ runs.
Conduct Experiments: Run the tests randomly to avoid systematic bias.
Analyze Main Effects: Calculate the average change in the output response when a factor goes from its low to high level.
Analyze Interaction Effects: Calculate how the effect of one factor changes depending on the level of another factor.

ANOVA in Design of Experiments

Analysis of Variance (ANOVA) is the primary statistical tool used to analyze the results of a DoE. While main effects plots can visually suggest which factors are important, ANOVA provides the mathematical rigor to prove it.

Partitioning Variance: ANOVA breaks down the total variance observed in the experimental results into specific sources: the variance caused by Factor A, the variance caused by Factor B, the variance caused by their interaction (A $\times$ B), and the unexplained variance (Error/Residuals).
Statistical Significance (p-values): By comparing the variance caused by a specific factor to the background error variance (calculating the F-statistic), ANOVA generates a p-value for every factor and interaction. If the p-value is less than the chosen significance level (e.g., $\le 0.05$ ), that factor or interaction has a statistically significant effect on the response.
Model Reduction: Once ANOVA identifies which factors and interactions are significant, researchers can build a predictive mathematical model using only those significant terms, ignoring the insignificant ones.

Advanced Optimization Techniques

When models become complex, civil engineers rely on more advanced DoE methodologies.

Fractional Factorial Design: Used when testing many factors (e.g., $5$ or more) where a full factorial design would require too many runs. It tests a strategically chosen fraction of all possible combinations, sacrificing some higher-order interaction information to save significant time and money.
Taguchi Methods: Developed by Genichi Taguchi, this approach focuses on "robust design"—making a product or process insensitive to environmental variations (noise). It heavily uses orthogonal arrays and emphasizes minimizing the variance of the output, not just hitting a target mean.

Interact with the simulation below to explore the Taguchi Loss Function and its focus on variance.

Taguchi Loss Function Simulator

Adjust the actual measured thickness of the bridge bearing. Notice how the loss increases quadratically as you deviate further from the ideal target value of $20\text{ mm}$ .

Measured Thickness (

y

) [mm]: 20.20

19.0 mmTarget: 20.0 mm21.0 mm

Calculation Parameters

Target Value ( $T$ ): 20.0 mm
Loss Coefficient ( $k$ ): $2000/mm²
Deviation ( $|y - T|$ ): 0.20 mm

Mathematical Model

L(y) = 2000(20.20 - 20)^2

Calculated Hidden Loss

$80.00

Cost incurred due to variation from the ideal target.

Response Surface Methodology (RSM)

Response Surface Methodology (RSM) is a collection of mathematical and statistical techniques used to model and analyze problems where a response of interest is influenced by several variables, and the objective is to optimize this response. It explores the relationships between the explanatory variables and one or more response variables, often using polynomial regression models to map a "topographical" surface of the data space to find the absolute peak (optimum) performance.

When conducting an RSM analysis in civil engineering (like optimizing a complex concrete mix design involving multiple new chemical admixtures), researchers generally employ one of two specific experimental designs to map the curvature of the response surface efficiently:

Central Composite Design (CCD): The most popular RSM design. It starts with a standard full or fractional factorial design (the "cube" points), adds several runs precisely at the center point (to estimate the inherent experimental error and curvature), and then adds "star" or "axial" points extending outwards from the center along each axis. This design is highly flexible and excellent at mapping the full quadratic surface.
Box-Behnken Design (BBD): An alternative RSM design specifically created to estimate quadratic models. Unlike CCD, BBD does not test points at the extreme vertices of the experimental region (where all factors are simultaneously at their highest or lowest levels). This is highly advantageous in engineering when these extreme combinations might be physically impossible, prohibitively expensive, or dangerously volatile to test in the laboratory.

Key Takeaways

DoE is far more efficient and informative than changing One-Factor-At-A-Time (OFAT), allowing engineers to discover critical interaction effects between multiple variables and optimize complex engineering processes and material formulations.
Full factorial designs test all possible combinations of factor levels (e.g., a $2^k$ design).
ANOVA mathematically partitions the total experimental variation to pinpoint exactly which factors and interactions are truly driving the results. It uses p-values to separate statistically significant effects from random noise and is required to build accurate predictive models.
Fractional factorial designs save resources when screening many factors by strategically testing a subset of all possible combinations. Taguchi methods focus on making processes robust against uncontrollable noise.
Response Surface Methodology (RSM) uses regression modeling to find the absolute optimal settings in a multi-variable space.
Central Composite Design (CCD) is the standard RSM tool that adds center and axial points to a factorial design to map out the entire curvature of the response surface.
Box-Behnken Design (BBD) is an RSM design that estimates quadratic models without testing the extreme high/low combinations of all factors simultaneously, making it ideal when those extreme points are physically impossible or unsafe to test in the laboratory.

PreviousWriting and Presenting Research - Examples & Applications

Quiz Me

NextDesign of Experiments and Optimization - Examples & Applications