DOE fundamentals -
under six minutes a lesson
Eight short lessons on Design of Experiments, written for bench scientists - no statistical background assumed, no jargon kept for its own sake. Read them here, get them by email, or watch the video masterclass.
Dear Scientists,
I spent more than a decade at the bench before anyone explained DOE to me in a way that survived contact with a real plate map. The courses I found assumed I wanted to become a statistician. I did not - I wanted my experiments to stop wasting my time.
These eight lessons are the course I wish someone had handed me: what multifactorial design actually buys you, how to frame a design space, when to screen and when to optimize, and how many runs an honest answer costs. The last lesson shows the whole method applied in Shadow AI - because the fastest way to learn DOE is to run one.
Ankita Pandey
Founder and CEO, Shadow AI
Eight lessons, in the order the method works
- 01WHY DOE
One factor at a time is the slowest way to learn
Most labs still vary one factor at a time - OFAT. It feels careful, but it is statistically the weakest and most expensive way to explore a system. This module shows why multifactorial designs find interactions OFAT is structurally blind to, with fewer runs.
Watch module 01 ↗ - 02THE DESIGN SPACE
Factors, levels, responses - your experiment as a design space
Before any design can be chosen, the question has to be framed: what can you turn (factors), how far can you turn it (levels and ranges), and what will you measure (responses). Getting this frame right is most of the work - and most of what goes wrong.
Watch module 02 ↗ - 03SCREENING
Screening designs - find the vital few
With 6-10 candidate factors you do not optimize - you screen. Plackett-Burman and fractional factorial designs rank main effects in remarkably few runs, so the expensive follow-up work is spent only on factors that matter.
Watch module 03 ↗ - 04OPTIMIZATION
Response surface methods - find where it is best
Once screening has cut the field to 2-4 factors, response surface designs - central composite, Box-Behnken - map curvature and locate optima. This is where DOE stops saving runs and starts finding conditions you would not have tried.
Watch module 04 ↗ - 05RIGOR
Replication, randomization, blocking - the three defenses against noise
Biology is noisy; the design has to defend itself. Replication estimates noise, randomization protects against drift and confounding, blocking removes the variation you already know about - day, plate, batch, operator.
Watch module 05 ↗ - 06STATISTICAL POWER
Power and sample size - how many runs you actually need
Underpowered experiments are the quiet failure mode of bench science - they cannot see the effect they were built to find, and the null result gets believed anyway. This module makes the effect size / noise / run count trade-off concrete.
Watch module 06 ↗ - 07THE WALKTHROUGH
From research question to DOE in Shadow AI
The product module. A real research question goes in; Shadow AI runs the literature across PubMed, OpenAlex, Semantic Scholar and arXiv, proposes ranked hypotheses, and returns a bench-ready design - factors, levels, controls, materials, statistical plan - in minutes. Everything from modules 01-06, applied.
Watch module 07 ↗ - 08ANALYSIS & ITERATION
Reading the results - and designing the next campaign
A DOE is rarely one experiment - it is a campaign. This module covers what to look at first in the analysis (effects, then model fit, then residuals), what a 'failed' design still teaches you, and how the result of one design seeds the next.
Watch module 08 ↗
One lesson a day, eight days
The same eight lessons, delivered one per day - each readable in under six minutes, ending with the Shadow AI walkthrough.
Lesson 09 is your own experiment
Shadow AI turns a plain-language research question into a bench-ready design - hypotheses, DOE, controls, materials, statistical plan. The Explorer plan is free.