New: building the AgentXRD expert workbench at the OSI hackathon, MIT, October 21 to 22 →
Environments for materials AI

Teaching AI the judgment of materials engineering.

We turn messy lab and engineering data into scientific environments where your agents use real tools, check their work against experimental evidence and learn from expert corrections. The goal is agents that test competing explanations, revise their analyses and improve on new cases. We start with X-ray diffraction.

Your data, your agents, your infrastructure. You decide what gets used for training.

See AgentXRD
First workbench · AgentXRD
Useful today. Learning for tomorrow.
A good fit is a checkpoint, not a conclusion. We're building AgentXRD so scientists and agents can look into features a fit leaves unexplained, compare other interpretations, add expert corrections and rerun the analysis. Those decisions and their outcomes become material for evaluating agents and, with your permission, training them. The plot below shows the first step, the agent's own fit.
AgentXRD · Rietveld fit · Rwp 3.54% · GOF 1.8 NiO 37.9 · Bi₂O₃ 22.3 · Li₂CO₃ 39.8 wt%
A powder mixture from the public DARA benchmark. An agent running on a local open-weight model refined it in BGMN. It found the same three phases as the published refinement, at close to the same weight fractions. White dots are the measurement, the cyan line is the agent's fit, and the gray trace underneath is what the fit misses.

There is no standard Rwp cutoff for a good fit. Crystallographers compare the misfit with what counting noise alone would leave. That ratio is the goodness of fit (GOF), and 1 is ideal. They also check that the difference curve is flat and that the phases make chemical sense (Toby, 2006). Here the misfit is 1.8 times the noise.

The feedback loop
Scientific work that checks itself and teaches the next agent.
Useful scientific AI has to do more than produce an answer. It has to test that answer against the measurements, state its assumptions, look into disagreements, and ask an expert or request another measurement when the evidence can't decide. We're building those checks into the workflow, so each success, failure and correction becomes feedback the next agent can learn from.

Self-verifying feedback loops. Agents check their work with scientific tools and experimental evidence. Questions the evidence can't settle go to experts or to further experiments.

  1. InvestigateScientists and agents analyze real measurements with the lab's own tools.
  2. Check against evidenceTool results and the measurements themselves test the answer.
  3. Revise or seek helpAn unexplained result triggers another analysis, an expert review or a new measurement.
  4. Learn with permissionTraining on your data, or sharing what it teaches, happens only with your explicit permission.
  5. Test on new casesThe next agent is checked on cases it hasn't seen before anyone relies on it.

Automate what can be checked. Surface what needs judgment. Measure whether the next agent actually improves.

Where we're going
AI that engineers materials as reliably as the best materials engineers.
That takes real experiments from many labs, not just one.

Work with us.

For teams training materials AI, diffraction experts, labs with hard XRD patterns, and engineers and scientists who want reliable AI agents.

Book a 30-minute call

Or email gabe@detrix.ai.