Training models to discover in the real world.

Iterative Labs is building training environments and self-driving laboratories for AI that learns from real experiments, not only simulated ones.

Become a contributing scientist Backed by Y Combinator

Fig. 0 — A self-driving bench. Liquid handler, transfer arms, plate reader, incubator. Step transfer

01 / Thesis

Simulation is where a model searches. The physical world is where it is graded.

Text records conclusions, not the search that produced them. Simulation restores the search, but a model trained inside one learns the simulator, not the world.

The bottleneck is not intelligence. Models can reason about science on paper, but they have never had to learn from what happens at the bench. We build the environments and the laboratories where they do.

02 / Environments

Domains where simulation is informative and experiment is decisive.

SimulatedMeasuredWhere they disagree

Biology

Binding affinity

Sim: Docking, free-energy methods
Exp: Plate assays, SPR

Chemistry

Reaction kinetics

Sim: DFT transition states
Exp: Measured rates and yields

Materials

Phase stability

Sim: DFT convex hull
Exp: Synthesis, X-ray diffraction

Physics

System dynamics

Sim: FEM, CFD
Exp: Instrumented bench data

03 / The iterative loop

A proposal becomes an experiment. A measurement becomes a reward.

The model explores in simulation. Search selects the candidates the simulator cannot settle. Those are synthesized, measured, and scored against what was predicted. The score trains the model; the discrepancy trains the simulator.

iSimulation 2,400 candidates scored

iiSearch 3 selected for synthesis

iiiExperiment Solid-state synthesis, Ar

ivMeasurement XRD · predicted vs. measured

vReward r = 0.71 · simulator updated

Fig. 1 — One cycle, shown for solid-state synthesis. State Simulation

04 / Self-driving labs

Agents that run the laboratory end to end.

Design an experiment, simulate it, write it as a protocol a machine can execute, run it, and make sense of noisy, corrupted, or failed results. Each environment teaches an agent that whole loop.

The physical side begins in biology: automated liquid-handling protocols such as cell-free protein expression, run on a rented wet-lab bench. Every real outcome is cached as ground truth that agents are trained and graded against. Chemistry and materials follow.

Protocol Written by the agent · validated before it runs

Run Automated liquid handling · biology first

Readout Failed and noisy runs included

Fig. 2 — A cell-free protein expression run, from protocol to readout.

05 / Sim-to-real

Not every problem belongs in the loop.

Simulation suffices

The experiment is redundant.

Orbital mechanics · Lattice constants

Where we work

Simulation narrows. Measurement decides.

Catalyst selectivity · Thin-film growth

Simulation fails

The experiment is a guess.

Polymer aging · Cell culture yield

Fig. 3 — Simulated against measured outcome, in three regimes.

06 / Contributing scientists

Know a scientific problem models should learn to solve?

We are interested in researchers and domain experts working on tasks that can be explored in simulation and verified experimentally.

Existing data: Measurements you have already taken, including the data behind your published papers.
Expertise: Your time defining tasks and reviewing results in your field.

Contributing scientists are compensated for both.

Company

Founders

CEO

Kajen Thavaraj

Built voice AI at Neuralink and robotics AI at Figure. Studied computer engineering at the University of Toronto.

LinkedIn kajenthavaraj@gmail.com

CTO

Tommy Lee

Trained ML models that translate thought into action at Neuralink. AI at NVIDIA, and crash-prevention models at Tesla. Studied computer engineering at the University of Waterloo.

LinkedIn tommy.lee@uwaterloo.ca

Neuralink NVIDIA Figure Tesla