Research automation grounded in scientific practice
Discovery Loop is more than a prediction model presented as an automation tool. Its loop architecture draws on active learning, Bayesian experimental design, and the way researchers organize real discovery projects.
Where Discovery Loop is used today
The engine is domain agnostic. Adapters and objective functions change with each field. These are the research settings where teams currently run the loop.
Materials science
Optimize synthesis conditions, sweep compositions for mechanical or optical properties, and test stability across temperature and humidity ranges.
Chemical synthesis
Optimize solvent, temperature, catalyst, and stoichiometry conditions. Maximize yield and purity using HPLC and NMR readouts.
Biochemistry and cell biology
Optimize assay conditions, characterize dose response curves, and map enzyme kinetics from plate reader data.
Formulation science
Select excipients and sweep ratios for pharmaceutical and consumer product formulations. Measure stability through accelerated testing.
Agricultural research
Optimize crop treatments, test soil amendments, and sweep growth conditions using controlled environment data.
Energy materials
Optimize electrode composition, sweep electrolyte formulations, and characterize cycle life from electrochemical test data.
Active learning, beyond prediction
Predictive models learn from historical data and estimate results under unseen conditions. They do not choose experiments. Discovery Loop uses active learning to ask which experiment would reduce uncertainty most efficiently if performed.
Active learning closes the decision loop. The model is not a static artifact for researchers to consult. It issues the next experimental instruction, creating the architectural basis for automation.
Hypothesis trees over flat trial logs
Many research automation tools represent experiments as an unstructured trial list. Discovery Loop organizes hypotheses as a tree, a branching probability space covering possible outcomes. Each experiment prunes a branch by ruling out a region or extends one by adding candidates for exploration.
The structure shows which parts of the search space remain open, rather than only listing tested conditions. It also gives researchers a human readable record of how the project developed.
Hypothesis tree snapshot, Cycle 14
Information gain per wetlab dollar
Reinforcement learning and bandit algorithms usually minimize expected regret, the difference between achieved and optimal outcomes. Scientific research has another cost structure. Instrument runs are expensive, reagents are limited, and time matters more than episode count.
Discovery Loop targets information gain per unit of run cost instead of expected regret. The policy considers run cost, instrument lead time, and reagent availability. It chooses experiments that advance the work efficiently under real constraints, rather than simply maximizing a theoretical reward signal.
Run cost weighting
Each instrument class carries a configurable cost weight. The loop deprioritizes expensive instruments when cheaper alternatives yield similar information gain.
Reagent constraints
Experiments needing reagents below the inventory threshold are automatically delayed until restocking, or excluded when the threshold is zero.
Lead time awareness
Scheduling includes instrument lead times and batch sizes. The loop avoids queuing a 48 hour run when an informative 2 hour alternative exists.
Read the technical notes
The team shares detailed technical writing about the loop's methods. Begin with the notes on hypothesis trees and the explore exploit tradeoff.