Chemicals & Materials

Materials discovery.
Fewer, better experiments.

Generative models can propose more candidate materials than any lab could ever make. The useful system is the one that decides which handful to make next, and learns from what happened.

All Chemicals & Materials →

What it is

An AI materials discovery pipeline combines machine-learning property models, physics-based simulation such as DFT where it is affordable, public data from sources like the Materials Project, and your own measurements, in an active-learning loop. Each round, it proposes the candidates whose results would most improve either performance or the model’s knowledge, your scientists choose what to synthesise, and the results feed the next round.

Machine learning in materials science is strongest where data is plentiful and weakest at the edge of the known, which is exactly where discovery happens. So we design for small data: uncertainty estimates on every prediction, transfer from public datasets where the chemistry is related, and an honest report when the model is extrapolating.

What decides whether it works

Your data, joined

Synthesis conditions, characterisation and performance results linked per sample. Without that record, no model can learn from your lab.

Honest uncertainty

Every prediction carries its uncertainty, and ranking uses it, so the loop explores where it is ignorant instead of over-trusting where it is confident.

Not: proof of a material

A predicted stable, high-performing material is a candidate until it is made, measured and reproduced. Publications and claims rest on your lab’s results, not the model’s.

A discovery round

ACTIVE-LEARNING LOOP

01

Design space

compositions, structures, process window
02

Predict

ML models and simulation, with uncertainty
03

Select

best expected value of the next experiment
04

Synthesise and test

scientists choose and run
05

Record

conditions and results, per sample
06 · Branch / escalation

Retrain

the next round starts better informed

Solid: one roundDashed: results feeding the next

The pipeline’s job is to make each synthesis count, not to replace synthesis.

Questions

Questions about Materials Discovery.

What is AI materials discovery?

It is the use of machine learning, often alongside simulation, to search for new materials with target properties faster than trial and error. Models learn from known materials and your measurements, predict which untested candidates are promising, and choose experiments that teach the most. The output is a shortlist for the lab, not a verified material.

How is machine learning used in materials science?

For property prediction from composition or structure, for speeding up simulation with learned potentials, for analysing characterisation data such as spectra and micrographs, and for steering experiments through active learning. Results are best where training data is close to the problem and least reliable when extrapolating, so uncertainty matters as much as the prediction.

Can a small R&D team use AI for materials discovery?

Yes, and active learning suits small teams because it squeezes more information out of each experiment. The prerequisites are a clear target, a defined design space and a consistent record of what was made and measured. Compute can be rented; the lab discipline cannot, so that is where we start.