Response Surface Methodology
Response surface methodology fits low-order polynomials to sampled data to map how outputs respond to inputs and to locate optima.
Origin and idea
Response surface methodology (RSM) grew out of industrial experiment design in the mid-twentieth century. It approximates an unknown response y = f(x) with a simple polynomial - usually linear or quadratic - fit by least squares to a set of designed experiments. The fitted surface is then used to understand trends and to move toward an optimum.
The quadratic model
A second-order response surface has the form y = b0 + sum(b_i x_i) + sum(b_ii x_i^2) + sum(b_ij x_i x_j). The linear terms capture main effects, the squared terms capture curvature, and the cross terms capture two-factor interactions. For k inputs this model has (k+1)(k+2)/2 coefficients, so the design must supply at least that many distinct runs.
Designs that feed RSM
- Central composite designs add axial and center points to a factorial base to estimate curvature
- Box-Behnken designs avoid extreme corner combinations
- Factorial and fractional-factorial designs estimate main effects and interactions economically
Fitting and diagnostics
Coefficients come from ordinary least squares. The fit is judged by the coefficient of determination, adjusted R-squared, lack-of-fit tests against replicated points, and residual plots. Because the model is deliberately simple, RSM is most reliable over a small, local region of the input space where a quadratic is a fair approximation of the true response.
Optimization by steepest ascent
When the current region is far from an optimum, the gradient of the fitted linear surface points the direction of steepest ascent. Experiments march along that path until improvement stalls, then a new local quadratic is fit to find the peak precisely. This sequential strategy is the classic RSM optimization loop.
Where RSM ends
A global quadratic cannot represent multi-modal or sharply nonlinear responses. For those, RSM gives way to richer surrogates: polynomial chaos, Gaussian processes, or neural networks. Still, RSM remains the fastest way to screen a handful of design variables and to get an interpretable first map of a response before committing to heavier machinery.