OpsightDefine the search.
Set the objective, available parameters and constraints. Bring in relevant campaign data, scientific context and human hypotheses.
A new direction for experimental design
Choose the next experiment.
With more insight.
Opsight explores how experimental data, scientific reasoning, optimization and human expertise can work together to help R&D teams decide what to test next and learn from every result.
Research-led optimization for experimental R&DFor teams exploring formulations, materials and processes.
01 The problem
More capacity doesn't always
mean more clarity.
Running more experiments doesn't guarantee a clearer answer. The difficult question isn't just how to run one, but which is worth running next.
Useful knowledge sits in past campaigns, papers and the experience of your team. Meanwhile, the space of possible experiments grows with every parameter you can change.
Bringing those pieces together can help guide experimental searches with both the data and the science behind it.
02 The approach
A loop that learns
from every experiment.
The research combines probabilistic optimization with hypothesis-led reasoning, keeping scientific judgment in the loop and recommendations open to review.
Choose what to test.
Use optimization and reasoning to propose experiments. Inspect the hypotheses and rationales before deciding what to run.
Learn and refine.
Return experimental results to the search. Update hypotheses, reconsider assumptions and use the new evidence in the next decision.
The result is more than a recommendation. It is a record of what you tried and what you learned.
03 Research, not just promise
Scientific reasoning
can improve the search.
Our published work compares Bayesian optimization, LLM-only reasoning and hybrid approaches, and examines their limitations, not just their best runs.
Published photocatalysis benchmark · 10 variables · 150 virtual experiments · 20 repeat runs · Batch size 10
BORA hybrid · o3 + Bayesian optimization
25.3 ± 2.9
µmol h⁻¹ · mean best hydrogen evolution rate
Tested Bayesian optimization baseline
11.6 ± 4.4
µmol h⁻¹ · mean best hydrogen evolution rate
Cissé et al. · Digital Discovery, 2026 · Table 3
Values are mean ± standard deviation across 20 runs.
These benchmark results are part of a broader research program spanning computational studies and noisy, closed-loop experimentation on a robotic laboratory workflow.
04 Work with us
Start with a real
optimization challenge.
We're looking for R&D partners to help validate where this approach could be useful.
Planning new experiments
Explore a closed-loop pilot.
Tell us what you want to improve, what you can vary and how you measure success. Together, we can discuss a scoped evaluation that fits your experimental constraints.
Discuss a pilotAlready have experimental data?
Explore a retrospective study.
Historical settings, outcomes and relevant constraints can provide a starting point. We'll discuss feasibility, data coverage and suitable comparisons before agreeing a study.
Discuss your dataset05 About Opsight
Finding an optimum
should also mean
finding new knowledge.
Built around how scientists work: propose, test, challenge and revise.
Opsight builds on research into hypothesis-led optimization and scientific reasoning in closed-loop experiments at the University of Liverpool.
Our aim is to complement domain expertise and laboratory workflows, not replace them. Recommendations and their supporting rationales should be inspectable, while the distinction between a hypothesis and a validated explanation stays clear.
Meet Dr Abdoulatif Cissé and Dr Miriam Wilson at ELRIG to discuss your problem, or get in touch below.
A few useful questions
Before we talk.
Do I need an automated laboratory?
Not at all. Tell us how you currently run experiments and collect results. We can discuss what a suitable automated or manual workflow would require, including any integration work.
Can I start with an existing dataset?
Yes. Describe it in the inquiry form. Useful datasets include the settings for each experiment, measured outcomes and relevant constraints. Feasibility depends on quality, coverage and noise, not just the number of rows.
Does the AI explain its recommendations?
Opsight's hypothesis-led approach records proposals and rationales for review. These are useful for scrutiny, but they do not reveal every model internal or establish that a suggested mechanism is correct before validation.
Can I try Opsight?
We're inviting R&D partners to explore scoped pilots and retrospective evaluations. Together, we'll assess fit, define what to evaluate and learn where the approach is useful.
06 Let's explore the fit
What are you
trying to optimize?
Tell us a little about your work and what you'd like to explore. You can request a demo, discuss an optimization challenge or explore an existing dataset. We'll get in touch using the details you provide.
No account or dataset upload required.
At ELRIG?
Find us at Breakthrough Zone BZ42.
Our scientific poster is 74 in the Exhibition area.