Citrine
Design of Experiment (DoE) or AI-guided experimentation?
When Does DoE Stop Being Enough?

Citrine has seen a 50%-90% reduction in the number of experiments needed to reach target performance.
Classical Design of Experiments, or DoE, is a proven method for studying a small, stable, well-defined design space.
But specialty chemicals and materials product development rarely stays that simple. Variables multiply. Raw materials change. Teams balance performance, cost, processing, and other constraints at once.
AI-guided experimentation helps scientists decide what to test next, using both predicted performance and estimated uncertainty.
“I’ve spent so much time trying to get our scientists to use DoE tools, and I’ve had much better adoption with Citrine. It goes to show how important the usability is.”
– Melissa Helm, R&D Manager, Process Chemistry & Technology, Stepan
Estimated uncertainty: how it improves the next experiment
A prediction answers one question: “Which candidate is expected to perform well?”
Estimated uncertainty answers another: “How confident is the model in that prediction?”
For each potential experiment, AI-guided experimentation estimates likely performance and the confidence associated with it. High uncertainty can show that an area of the design space is underexplored or that the team has limited relevant data.
This gives scientists more useful information when planning the next lab cycle. They can select experiments that:
- are predicted to meet the target;
- have a high probability of success; or
- could generate valuable learning by reducing uncertainty.
Classical DoE can quantify model fit and variation after experiments are run. AI-guided experimentation adds a prediction-specific estimate of uncertainty before the next experiments are selected.
That extra information helps teams balance progress toward a target with learning, making each experimental cycle more productive.
What Is AI-driven iterative experimentation?
AI-driven iterative experimentation (sometimes called Sequential Learning or Active Learning) is an iterative, AI-guided R&D methodology where each experiment<>modeling loop improves the AI model that is used to help select the next batch of experiments to perform. The AI model identifies promising candidates to synthesize and test in each round. Including uncertainty in each prediction is critical to get the most value out of this approach.

- Start with available experimental, simulation, or historical data.
- Define the target properties and scientific constraints.
- Review candidate experiments by predicted performance and estimated uncertainty.
- Choose the experiments that best fit the team’s objectives.
- Add the new results and use them to guide the next cycle.
Using the Citrine Platform you can model 26 properties as easily as 2 properties just by adding in some new columns to Excel.
Scientists remain in control throughout. The AI provides additional decision information. Scientists apply their expertise to decide what belongs in the lab. Empowering product experts with this extra information reduces the number of experiments needed to reach performance targets by 50-90%. You can read our Case Studies to see how this pans out in practice.
For a comparison of more advanced modeling and design approaches like Bayesian Optimization and Response Surface Analysis, read this blog.
Choose the method that fits the problem
Use classical DoE when the design space is compact, stable, and clearly defined, with a limited number of controllable variables.
Consider AI-guided experimentation when the work involves many variables, several target properties, changing constraints, or historical data from multiple projects or laboratories.
Here is how to decide between them.
Classical statistical DoE
AI-guided experimentation
Starting point
DoE
A defined experimental matrix. The number of experiments needed for DoE goes up exponentially with the number of dimensions to optimize.
AI-guided experimentation
Approximately 10–20 initial data points can be practical
Design space
DoE
Defined at the outset
AI-guided experimentation
Can be refined after any experiment
Variables
DoE
Practical limits emerge as variables multiply
AI-guided experimentation
Can incorporate many variables, including molecular descriptors
Historical data
DoE
Requires closely comparable runs
AI-guided experimentation
Can incorporate heterogeneous historical data, with uncertainty reflecting local confidence
Missing data
DoE
Can disrupt a fixed matrix
AI-guided experimentation
Can be accommodated, with uncertainty increasing where evidence is limited
Scientific and business constraints
DoE
Expressed through design choices and factors
AI-guided experimentation
Applied directly alongside property targets
Best fit
DoE
Small, well-defined, stable spaces
AI-guided experimentation
Complex, evolving, multi-variable development problems
When classical DOE is the right choice
DoE remains a strong option when a team has limited or no usable starting data, a small number of controllable variables, and a stable, well-defined design space. It can be especially useful for local optimization and for establishing an initial understanding of an unfamiliar process. Citrine’s customers often use a streamlined, high-dimensional DoE to kick off new projects where no relevant experimental data is available.
AI-guided experimental cycles become more useful when the team needs to account for many variables, reuse heterogeneous data, balance several target properties, or learn across repeated experimental rounds.
