Closing the loop: the experiment cycle bottleneck
Research slows most between receiving a result and choosing the next run. Learn how to quantify that delay and what automation changes.
Notes
Discovery Loop notes on automating experiments, active learning, and the science that guides the loop.
Building AI to read lab instrument output
Lab instruments emit vendor-specific formats that rarely agree. This account explains adapters for extracting dependable signals from HPLC, plate readers, and spectroscopy systems.
Hypothesis trees and AI-driven experimental design
We use a tree, not a flat trial queue, to map the search. See how branching hypotheses expands the information guiding the decision engine.
Materials optimization in 20 cycles: a case study
See how a materials lab used Discovery Loop to select the ideal synthesis setting for a coating formulation in 20 experiment cycles.
The explore-exploit tradeoff in automated science
The loop weighs probing unknown parameter regions against pursuing its strongest candidates; the proper mix is set by experiment cost.
Reading HPLC data: instrument integration notes
Agilent, Waters, and Shimadzu HPLC files encode identical information differently. Learn how each adapter processes them and how the standardized result appears.
Why science needs agentic AI, not predictive AI
Predictive AI estimates the outcome under given conditions; agentic AI selects the next experiment. Their distinction is architectural, not merely semantic.
Parallel experiment queues and scheduling
With several instruments available, the loop schedules parallel runs to increase information gained per unit time. Here is how it manages their queues.
Building the research loop: architecture notes
Explore the architecture of Discovery Loop: adapter layer, analysis module, decision engine, plus the channels connecting them.
Search-Space Limits: Why Labs Run Few Experiments
A lab's parameter space can be vast, while its annual experiment capacity is small by comparison. That gap constrains what discovery can reach.
Active learning for scientists: a literature review
Review active learning work most applicable to automating physical-sciences research: acquisition functions, Bayesian optimization, and uncertainty quantification.
Why we built an experiment orchestrator, not a prediction model
Discovery Loop was founded on a key choice: automate decisions rather than merely improve prediction. See how that principle directs the product.