故事 · 为什么不能一次只动一个
Origin Story · Why one-at-a-time fails
想象在浓雾里爬一座山,只准沿正东或正南方向走。你先朝东走到东西方向的最高处,再朝南走到南北方向的最高处 ——
如果山是正南正北对称的,这样能爬到顶。但只要山峰斜着长(因子之间有交互),你这种「一次一因子」的走法就会停在半山腰,
误以为到顶了。农学家 Ronald Fisher 早在 1920 年代的田间实验里就看穿了这点:
与其一块地只变一个变量,不如用正交设计把多个因子同时变,让每次收成都同时携带多个因子的信息。
这就是 DOE 的起点 —— 用更少的地、更少的实验,看清整片地形,而不是只沿两条轴瞎走。
Picture climbing a hill in dense fog with only two legal moves: due east or due south. Walk east to the highest east-west point, then walk south to the highest north-south point —
if the hill is perfectly aligned with the compass, you'll reach the top. But the moment the summit tilts off-axis (the factors interact),
that one-at-a-time march parks you on a shoulder of the hill and you swear you're at the peak. The agronomist Ronald Fisher saw through this back in the 1920s, in his field trials:
instead of changing one variable per plot, run an orthogonal design that varies several factors at once, so every harvest carries information about every factor.
That's the seed of DOE — fewer plots, fewer runs, and a clear view of the whole terrain instead of stumbling along two axes.
1 因子空间与响应曲面:那座斜着的山 Factor space and response surface: the tilted hill
拖滑块探索Drag to explore2 OFAT 与正交设计的根本区别 What separates OFAT from an orthogonal design
OFAT 一次一因子:固定其余、只动一个,画出十字布点。看不到交互,最优点常错过。
OFAT (one factor at a time): hold everything else fixed, move one knob, trace a cross-shaped layout. Interactions stay invisible and the true optimum slips by.
⊞ 全因子/正交:因子组合铺成网格,每点同时带多因子信息,主效应与交互一起解出。
⊞ Full-factorial / orthogonal: factor combinations form a grid, every run carries multi-factor information, and main effects plus interactions fall out together.
主效应:单个因子独立改变响应的平均效果,OFAT 也能估。
Main effect: the average shift in the response caused by changing a single factor on its own — OFAT can still estimate this.
交互效应:A 的最佳值随 B 变化 —— 只有同时变多因子才看得见,正是 OFAT 的盲区。
Interaction effect: the best A depends on the level of B — only visible when several factors move together, and exactly the blind spot of OFAT.
3 现实里的实验设计 DOE in the real world
工艺参数优化:注塑的温度/压力/保压时间,正交表几组实验就锁定窗口,省料省工时。
Process tuning: a small orthogonal array nails down the injection-molding window across temperature, pressure, and hold time — saving resin and hours of trial-and-error.
配方与反应:化工、制药的多组分配比,DOE 找出协同与拮抗,远胜逐个试。
Formulation and reaction: in chemicals and pharma, DOE reveals synergies and antagonisms between ingredients far faster than testing them one by one.
交互作用挖掘:催化剂只在高温下才起效 —— 这种「条件性最优」唯有 DOE 能捕捉。
Hunting interactions: a catalyst that only fires at high temperature — that kind of conditional optimum only shows up in a properly designed experiment.
响应曲面法(RSM):在最优区附近加密布点拟合曲面,精修到真正的峰顶。
Response Surface Methodology: densify runs near the optimum, fit a smooth surface, and polish your way onto the actual peak.
一句话In one line
DOE 的真正威力,不是「实验做得多」,而是每一次实验都同时回答多个问题。
OFAT 把因子拆开一个个问,看似严谨,却天生看不见交互 ——
而现实世界里,最优往往就藏在「A 和 B 一起调到某个组合」的斜坡上。
正交设计用结构化的布点,让有限的几次实验铺满整个因子空间,
一次性解出谁是主因、谁和谁纠缠,再用响应曲面把整张地形画出来,直奔峰顶。
这正是 DFSS 在设计阶段就敢拍板参数的底气:不是凭经验猜,而是用最少的实验把因果地图测出来。
The real power of DOE isn't "more experiments" — it's that every single run answers several questions at once.
OFAT interrogates one factor at a time, looks rigorous, and is structurally blind to interactions —
but in the real world the optimum usually sits on a tilted slope where A and B have to move together.
An orthogonal design uses a structured layout to blanket the whole factor space with just a handful of runs,
resolves which factor matters most and which pairs are entangled, and lets a response surface sketch the entire terrain — straight to the peak.
That's why DFSS dares to lock in parameters at the design stage: not by guessing from experience, but by mapping cause and effect with the fewest possible runs.
常见误用Common mistakes
沿用 OFAT 还自称做了实验。一次一因子抓不到交互,怀疑有交互时必须用正交/全因子。
Calling an OFAT campaign "a designed experiment". One-at-a-time can't see interactions — the moment you suspect one, switch to an orthogonal or full-factorial design.
因子和水平一把塞太多。先用筛选设计找关键少数因子,再精修,别让实验次数爆炸。
Cramming in too many factors and levels at once. Use a screening design first to find the vital few factors, then optimize — don't let the run count explode.
把局部峰当全局最优就收工。在最优区用响应曲面法加密验证,确认不是半山腰的假顶。
Declaring victory on a local peak. Densify runs in the optimum region with RSM to confirm you're at the true summit, not a false shoulder.