1
Forms hypotheses
Proposes what to try next
Turns a goal into a testable prediction about a material or physical system — a specific claim that an experiment can confirm or kill.
RAFATI · Los Angeles · The AI Scientist · Est. 2026
WHEREIDEAS
MEETREALITY
An AI scientist — and the autonomous laboratories where it learns to discover. Because knowledge isn’t read; it’s made.
Ideas in · Knowledge out
Building the AI scientist — and the labs where it learns.
RAFATI pairs AI models with automated experimental facilities so that software can form hypotheses, run real experiments, and learn from the results. Intelligence alone doesn’t move science forward — new knowledge is created only when an idea is tested against physical reality. So we give our models a place to act.
Our premise
Intelligence alone doesn’t move science.
Earlier scientific AI advances leaned almost entirely on models trained on internet text — a finite resource the leading models have already largely exhausted. The next breakthroughs won’t be read off the web. They have to be discovered.
New knowledge is created only when an idea is tested against the physical world. That is why we don’t just build smarter models; we build the autonomous laboratories where those models can act — on real instruments, real materials, and real results.
- Pair AI models with automated experimental facilities.
- Let the systems run experiments, not just reason about them.
- Make nature itself the feedback environment for learning.
2
Designs experiments
Plans the test
Chooses the conditions, controls, and measurements that will actually settle the question — so a run produces a clear answer, not noise.
3
Runs them autonomously
Acts in the world
Drives instruments in the lab to carry out the plan — no queue, no waiting on a free pair of hands. The system experiments by itself.
4
Learns from results
Updates its beliefs
Folds every measurement back into the model, sharpening the next hypothesis. The system gets better at the science with each cycle.
5
Keeps the negative results
Values every outcome
Records the experiments that “didn’t work” — the ones rarely published — because they tell the model exactly where not to look next.
6
Makes data that exists nowhere else
Builds the missing dataset
Every run yields high-quality experimental data found in no paper, dataset, or archive — the truth the internet never contained.
7
Works at the gigabyte scale
Measures deeply
A single experiment can return gigabytes of signal. The system reads all of it, finding structure no human could comb through by hand.
8
Simulates before it builds
Models first
Uses simulation to narrow the search before committing real lab time — one of the reasons physics is such fertile ground for this approach.
9
Closes the loop
Iterates fast
Hypothesis → experiment → result → revision, around and around — with nature as the judge and every turn faster than the last.
Why it matters
Why autonomous labs
Autonomous laboratories are the center of our strategy, not a convenience. They are how we get past the wall that scientific AI has run into — and how we build the data that the next generation of discovery will run on.
The internet ran out
Leading models have largely exhausted what the web’s text can teach. Progress now needs a new, deeper source of truth.
Data found nowhere else
Each facility produces large volumes of high-quality experimental data that exists in no dataset, paper, or archive on Earth.
Negative results, finally kept
The experiments that fail are seldom published — yet they’re some of the most valuable signal. We keep every one.
Tools to act, not just predict
Labs give the models real instruments. They run experiments in the world instead of only reasoning about them on paper.
Gigabytes per experiment
Every run is dense with signal — a single experiment can yield gigabytes of data for the model to learn from.
Built for scale
Automated facilities run continuously, compounding the dataset day after day — the engine quietly getting bigger.
The loop
How discovery happens
One full turn of RAFATI’s discovery engine, in three movements. The model designs; the lab runs; reality answers — and the answer becomes the next, sharper question.
01 Hypothesize & Design
01Set the goal. Pick a property worth chasing — say, a superconductor that works at a higher temperature than today’s materials.
02Form a hypothesis. Predict which materials or conditions might deliver it. Model
Simulation first
03Screen in simulation. Model candidates to narrow the search before touching hardware. Simulation
04Design the experiment. Choose conditions, controls, and exactly what to measure. Plan
02 Run & Observe
05Hand off to the lab. The plan becomes instructions for real instruments. Autonomous
06Run continuously. Automated facilities execute around the clock, no queue. 24/7
07Measure deeply. Capture gigabytes of signal from a single experiment. Gigabytes
Keep everything
08Record every outcome. Successes and the negative results others discard. Negatives kept
03 Learn & Iterate
09Check against reality. Physics is verifiable — results can be confirmed, the way math and code can. Verifiable
10Update the model. Fold the new data back in to sharpen the next hypothesis. Learn
Close the loop
11Iterate. A better hypothesis, a better experiment — faster than the turn before. Iterate
12Compound. The dataset grows with every cycle, holding signal found nowhere else. Compounds
Where this leads
What automated discovery unlocks
We’re starting in the physical sciences — high signal-to-noise, fast, simulatable, and verifiable. From there, automating materials design reaches into some of the hardest, most consequential problems there are.
Higher-temperature superconductors
Materials that carry current with minimal loss, closer to everyday temperatures.
Power grids with minimal losses
Transmission that wastes far less of what it carries — from plant to plug.
Cleaner transportation
Lighter, lower-loss systems for moving people and goods more efficiently.
Better semiconductors
Including the chip heat-dissipation problems that hold today’s hardware back.
Nuclear fusion
Materials and designs that help bring practical fusion power within reach.
Space travel
Materials engineered for the punishing demands of leaving Earth.
Materials design, automated
The engine beneath all of the above — discovery itself, sped up.
Deployed with industry
Custom agents that help partners’ engineers interpret data and iterate faster.
The labs & the programs
Where it all happens
The work lives in real places: the autonomous laboratory where experiments run, the simulation that scouts ahead of it, the industry floor where the methods meet production, and the programs that widen the circle of people pushing the same frontier.
Featured
The autonomous laboratory
Robotic handling, precision sensing, and the model, joined in one continuous loop. This is where a hypothesis becomes an experiment, an experiment becomes gigabytes of data, and that data becomes the next, better hypothesis — with physical reality as the only judge that counts.
Closed-loop · running continuouslyThe simulation environment
Scouts ahead of the bench
Where candidates are modeled before they’re ever made — narrowing the search and saving real lab time. Physics can be partly modeled, so the obvious dead ends are ruled out before a single sample is run.
The industry lab
Methods, meet production
Working alongside a semiconductor manufacturer on chip heat dissipation, we train custom agents that help its engineers and researchers interpret experimental data and iterate faster — the approach, applied where it earns its keep.
The grant program
Widening the circle
We’re launching a grant program to support outside researchers working at the same frontier — because the faster the whole field can test ideas against reality, the faster everyone discovers.
The scientific advisory board
Academic guidance
An academic scientific advisory board helps keep the science honest and ambitious — pairing fast-moving engineering with the depth of the research community.
Say hello
Let’s discover something together
RAFATI is built by people who have co-created widely known AI systems, contributed foundational machine-learning techniques and materials-science models, scaled autonomous physics laboratories, and taken part in real materials discoveries — backed by prominent venture investors and individuals for the long haul. If you’re a scientist, an engineer, an industry partner, or a builder who wants ideas to meet reality, we’d love to hear from you.
Thanks — your message is on its way
We’ve opened your email app with everything filled in. If nothing appeared, reach us directly at hello@rafati.co and we’ll get right back to you.