起源故事 · Box 与 Wilson 的化工厂 Origin Story · Box and Wilson at the Chemical Plant
1951 年,统计学家 George Box 和化学家 K.B. Wilson 在帝国化学公司研究如何最大化化学反应产率。 他们意识到:与其反复跑实验碰运气,不如把产率看成因子的连续曲面,用二次多项式近似它的形状, 再沿着最陡上升方向有方向地试验 —— 一步步逼近山顶。这套"响应面方法(RSM)"成了工艺优化的标准武器,至今仍是 DOE 的高阶招式。 In 1951, the statistician George Box and chemist K.B. Wilson were working at Imperial Chemical Industries on maximizing reaction yield. Their insight: stop firing off trials and hoping — treat yield as a continuous surface over the factors, fit a second-order polynomial to its shape, and then run experiments up the steepest-ascent path, marching toward the peak. That toolkit — the Box-Wilson response surface methodology (RSM) — became the standard for process optimization, and it's still the high-end move in DOE today.

1 响应曲面:一张等高线地图 The Response Surface as a Contour Map

爬山中climbing

2 切一刀:二次曲线就是一座倒扣的碗 Take a Slice: The Quadratic Is an Upside-Down Bowl

固定 X₂ 在你当前位置,沿 X₁ 切一刀,曲面就成了一条开口向下的抛物线。这正是"二次模型"的样子 —— 有弯曲、有极值。线性模型(直线)永远找不到这个顶点。 Fix X₂ at your current value and slice the surface along X₁ — you get a downward-opening parabola. That curve, with a turning point, is exactly what a quadratic model captures. A linear model (a straight line) can never locate this peak.

3 现实里的响应面 Response Surfaces in the Real World

产量最大化:化工反应的温度×浓度曲面,找出产率峰顶,就是 Box 当年的本行。 Maximizing yield: the temperature × concentration surface of a chemical reaction — find the yield peak. That was Box's day job.
成本最小化:把成本当"倒山",找谷底 —— 同样是二次曲面的极值问题。 Minimizing cost: treat cost as an inverted hill and chase the valley — same quadratic extremum, just flipped.
工艺窗口:等高线圈出"响应足够好"的稳健区域,告诉你参数能漂多远还合格。 Operating window: contour lines enclose the "good-enough" region — they tell you how far the setpoint can drift before quality leaves the door.
配方精调:在筛选后剩下的两三个关键因子上做 RSM,把配方拧到最优那一档。 Formulation fine-tuning: run RSM on the two or three vital factors that survived screening to nail the recipe to its sweet spot.
一句话In One Line
2ᵏ 全因子是侦察:它只在角上取样,告诉你哪些因子重要、大致往哪走。RSM 是精修: 加上中心点和星号点,拟出带弯曲的二次曲面,真正定位到峰顶。两者是接力 —— 先用最陡上升从山脚冲到山腰, 等斜率变缓(接近顶部),再补一组点拟合曲面、解出极值。别一上来就拟二次:离峰太远时,曲面只是局部近似,会把你带偏。 The 2ᵏ factorial is the scout — it samples corners, points out which factors matter, and shows roughly which way to head. RSM is the finisher: add center points and axial (star) points, fit a curved second-order surface, and pinpoint the actual peak. The two work in relay — use steepest ascent to sprint from the foothills to the upper slopes, and once the gradient flattens (you're near the top), bolt on a central composite design and solve for the stationary point. Don't fit a quadratic right away: far from the peak the surface is only a local approximation, and it will steer you wrong.
常见误用Common Mistakes
用线性模型找最优。极值在曲面内部,必须二次模型才有顶点,直线只会一路往边界跑。 Hunting for the optimum with a linear model. Interior optima need curvature. Only a quadratic has a stationary point — a plane just runs you to the boundary.
离峰很远就拟二次曲面。先用最陡上升靠近,斜率变平后再上 RSM 精修。 Fitting a quadratic far from the optimum. Use steepest ascent to get close first, then bring in RSM once the slope flattens.
外推到试验区间之外。曲面只在采样范围内可信,盲目外推峰顶可能根本不存在。 Extrapolating beyond the design region. The fitted surface is only trustworthy inside the sampled range. Extrapolated "peaks" may not exist at all.

响应面方法