A reactor block runs 24 experiments at once. A night shift can run a plate. The tempting plan is to fill every round. But every round is also the moment where results come back and the next experiments are chosen. Fewer rounds save calendar time; larger rounds give the planning fewer chances to learn. We measured where the balance lies.
Two simulations, one question
Study 1, reaction model.
540 simulated campaigns on a reaction model with 4, 8 and 12 factors. A start design, then rounds of 1, 2, 4, 8, 12 or 24 experiments until 96 experiments were spent. 30 paired instances per setting.
Study 2, public screens.
1,350 simulated campaigns on three published reaction screens in which every condition was measured. Budget 60 experiments: 5, 10 or 20 random starting runs, then rounds of 5, 10 or 20.
Per experiment, small and medium rounds are almost equal
After 96 experiments, planning eight per round instead of one at a time changed the final quality by 0.1 percentage points with 4 factors, not at all with 8 and by 1.1 with 12. Planning one experiment at a time is not worth the calendar time it costs.
Per round, larger is faster
With 8 factors, reaching 90 % of the best yield took:
| Experiments per round | Rounds | Experiments |
|---|---|---|
| 1 | 28 | 38 |
| 4 | 6 | 34 |
| 8 | 4 | 40 |
| 24 | 2 | 38 |
Same target, 4 rounds instead of 28, for two more experiments.
There is a ceiling
Very large rounds cost quality. After 48 experiments, rounds of 24 were 3.2 percentage points behind rounds of four with 4 factors and 2.2 behind with 12 factors. Rounds of 24 gave the lowest final quality at every factor count.
The public screens agree. With a budget of 60 experiments, 10 starting runs followed by rounds of 5 found on average 4.7 top-1 % conditions. The same start with rounds of 20 found 3.1, and rounds of 20 came last for every start size on every screen. Random picking found 0.6.
What we recommend
- One plate of material: do not spend it in one or two large rounds. Use a space-filling start, then rounds of 8 to 12 wells.
- Plates are cheap and time is short: full rounds of 24 are acceptable. They cost 1 to 3 percentage points but reach 90 % in 2 rounds.
- Large categorical screens: about 10 starting runs, then rounds of 5.
What the study does not show
Simulated campaigns on a model and on public screens with categorical factors. Real rounds also carry analysis delays and material constraints that the simulation does not model.
The full study
The 24-page whitepaper puts this into context: when classical DoE wins, when an adaptive approach needs fewer experiments, case studies and a decision guide.
FAQ
Why do larger rounds lose quality?
Every round is a chance to re-plan from all results so far. With a fixed budget, rounds of 20 or 24 give the planning only two or three such chances.
How many starting runs should I plan?
On the public screens, about 10 random starting runs followed by rounds of 5 was the strongest schedule. With 20 starting runs, the first 20 experiments are no better than random.
Does this apply to classical DoE?
A classical design is planned in one block, so round size does not apply in the same way. The results here are for adaptive campaigns, where each round is planned from the results before it.