Use case · Engineering & R&D

ML-guided R&D experimentation

Machine learning chooses which experiments, formulations or process recipes to run next, often on automated lab equipment, so R&D teams reach a working recipe in fewer rounds.

Maturity
Emerging
Complexity
High (our assessment)
Department
Engineering & R&D
Technology
Machine learning, Optimization & planning, Robotics & autonomy
Solves
Long lead times, Scrap & rework, Retiring expertise

Evidence at a glance

1documented deployment
1manufacturer
1independently assessed
Independently assessedManufacturer-reportedVendor-reported

Reported results

1 quantified outcome across 1 deployment

Independently assessedManufacturer-reportedVendor-reported

Each dot is one reported improvement, as published by its source; the black tick is the median. Figures come from different baselines and methods, so read them as the range of reported results, not a forecast.

Who has done it

Zymergen · Emeryville bio-engineering site

Machine learning directs the design-build-test-analyse-learn cycle so thousands of microbial strain experiments run at once on automated lines.

Independently assessed
Emeryville, CA · 2020
resulting in a doubling of its innovation rate
View the data as a table
ManufacturerEvidenceOutcome as reported
ZymergenIndependently assessedresulting in a doubling of its innovation rate

What it is

ML-guided experimentation uses machine learning, commonly Bayesian optimization and active learning, to decide which experiments to run next in product and process development. Instead of testing a fixed grid of formulations or process settings, the model learns from each result, estimates where better outcomes are likely and where it is most uncertain, and proposes the next batch of experiments that is expected to teach the most or improve the target the fastest. In some labs the proposals run directly on automated equipment such as liquid handlers, reactors, coating lines or test rigs, creating a closed loop of design, run, measure and update.

It matters because developing a new formulation, material, battery chemistry, semiconductor recipe or process window often means searching a space with many interacting variables, where each experiment is slow or expensive. Traditional one-factor-at-a-time work and even classical design of experiments can need many rounds to find a good region. Choosing experiments more intelligently can cut the number of runs needed, and keeping every result in a structured model preserves what was learned for the next project.

How it works

  1. Define the goal and the space

    Scientists and engineers set the targets, such as strength, purity, yield, cost or cycle time, the constraints that must be met, and the variables the experiments can change, including ingredient ranges, temperatures, times and process settings.

  2. Seed with existing data

    Past experiments, lab notebooks, literature values and a small initial design give the model a starting picture. Results must be recorded in a structured form with the conditions that produced them.

  3. Propose the next experiments

    The model estimates outcomes and uncertainty across the space and proposes a batch of experiments that balances exploring unknown regions with refining promising ones, while respecting safety and feasibility limits.

  4. Run and measure

    Experiments are run by technicians or on automated lab equipment, and measurements flow back into the data set, ideally directly from instruments and the lab information system rather than by hand.

  5. Update, decide and scale

    The model updates after each batch, and the team decides when results are good enough to move to pilot or production scale, where the same approach can tune the process window on larger equipment.

Benefits

Fewer experiments to reach a target

Choosing each batch based on what has been learned so far avoids running experiments in regions that are already understood or clearly poor, which shortens development cycles.

Better handling of interacting variables

Models capture interactions between ingredients and process settings that one-factor-at-a-time testing misses, which helps find combinations people might not try.

More use from lab automation

Automated equipment can run around the clock when the next experiments are chosen by software, instead of waiting for a scientist to analyze results and plan the next round.

Results that carry forward

Every experiment is stored with its conditions and outcomes in a structured model, so the next product or project starts from accumulated knowledge rather than scattered notebooks.

Data & prerequisites

Data you will need

  • Historical experiment records with conditions, ingredients and measured outcomes
  • Target properties, specifications and constraints for the product or process
  • Instrument and test results from lab systems such as LIMS or ELN
  • Raw material properties and lot information
  • Safety limits and equipment operating ranges
  • Pilot and production data when scaling a recipe

Before you start

  • Measurements that are repeatable enough for the model to learn real differences
  • Structured recording of experiment conditions and results, not only free-text notebooks
  • Scientists willing to run experiments the model proposes, including some that seem counterintuitive
  • A clearly defined objective, or an agreed way to trade off several objectives

Common pitfalls

  • Noisy or inconsistent measurements that make the model chase random variation
  • Defining the search space too narrowly, so the model can only confirm what the team already believed
  • Ignoring scale-up, so a lab-optimal recipe fails on pilot or production equipment
  • Buying lab automation before the data structure and decision loop are in place

Questions to ask vendors

  1. Which optimization methods do you use, and how do you handle several competing objectives and constraints?
  2. How does your system use our historical experiment data, including incomplete or inconsistent records?
  3. Which lab instruments, LIMS and ELN systems do you connect to?
  4. How do scientists review, override or add their own experiments to the model's proposals?
  5. Who owns the experimental data and models if we stop using your platform?

What to measure

  • Number of experiments or iterations to reach target specifications
  • Development cycle time from project start to approved recipe
  • Share of lab-developed recipes that succeed at pilot or production scale
  • Lab equipment utilization
  • Material and lab cost per development project

Frequently asked questions

How is this different from traditional design of experiments?

Classical design of experiments plans a set of runs up front and analyzes them together. ML-guided experimentation is sequential: it updates its model after each batch and chooses the next experiments based on what it has learned. The two are often combined, with a classical design used to seed the model.

Do we need a robotic self-driving lab to use this?

No. Many teams start with technicians running the experiments the model proposes. Automation adds speed and consistency once the loop is working, but the main gains come from choosing better experiments and recording results in structured form.

Can it be used on production processes, not only in the lab?

Yes. The same sequential approach can tune process settings on pilot lines or production equipment, where each trial run is expensive. In production, trials are usually constrained more tightly to protect quality and safety, and the work overlaps with process optimization.

Related use cases

Operations & production

Process & yield optimization

Models recommend or automatically adjust process setpoints to keep quality on target while improving yield, throughput and energy use.

Machine learningOptimization & planningIndustrial IoT & edge
34 documented deployments7 software companies

Quality

Quality prediction & virtual metrology

Predict product quality from live process data so problems are caught before parts are measured, tested or shipped.

Machine learningIndustrial IoT & edge
13 documented deployments2 software companies

Engineering & R&D

Generative design & engineering simulation

AI generates design options from requirements and replaces slow simulations with fast surrogate models so engineers explore more designs.

Machine learningGenerative AI & LLMsOptimization & planning
12 documented deployments3 software companies

Engineering & R&D

AI consumer insights for product development

Language models and machine learning mine product reviews, social posts, support tickets and sales data to show what customers want, so product teams pick stronger concepts and reach launch sooner.

Generative AI & LLMsMachine learning
1 documented deployment0 software companies

Put AI to work

Find the AI opportunities hiding in your plant.

Genvo, our implementation partner, works with manufacturers to pick the right first project, connect the systems you already run and put AI agents and tools into daily use, with people approving what goes out.

Assessments and consultations are provided by Genvo. How our partnership works.