Targeted Materials Prediction
Compute in advance the material, and how to make it, for a desired property. Accuracy and synthesis are the sticking points.
Open in the interactive tree →In principle quantum mechanics lets us compute any material property, but for real structures (defects, surfaces, temperature) calculations are too inaccurate or too expensive. Density functional theory (1960s) is the standard, but its approximations limit it. AI models learn from its results and inherit its errors.
As of October 2026
Machine-learning models predict the stability and simple properties of crystals well (GNoME: about 380,000 stable candidates, MatterGen: design to target values) but not reliably whether a substance can be made, how complex properties (superconductivity, catalysis, durability) behave, or which recipe leads there. Critics found that several 'new' A-Lab compounds were already known, Nature's January 2026 correction narrowed that novelty claim, and a June 2026 study finds that computed proposals depart from the structural patterns of experimentally discovered materials.
What is missing
- Calculation methods that describe strong electron correlation accurately and cheaply, for example quantum computers or better functionals
- Large, clean experimental datasets, including failures
- Prediction of synthesis routes and reaction kinetics
- Automated labs that test predictions at the same pace
Becomes possible once solved
- Materials designed on a computer instead of by trial and error
- Faster discovery of batteries, catalysts and superconductors
- Shorter development cycles in industry
Open steps
- Functionals for correlated electrons High AI leverageGet density-functional accuracy for strongly correlated materials, surfaces and defects at low cost, the weak point of today's calculations.
- Potentials for defects and surfaces High AI leverageMake machine-learned interatomic potentials accurate for defects, surfaces and finite temperature, not only ideal bulk crystals.
- Predicting how to make a material Medium AI leveragePredict precursors, temperatures and reaction pathways for a target crystal, not only whether it is stable.
- Self-driving labs and failure data Medium AI leverageAutomate synthesis, diffraction and property tests to validate predictions at the pace they are made, and record failed attempts in open datasets.
- Novel and makeable generated crystals Medium AI leverageMake generative models propose structures that are new to science and structurally like the materials that can actually be made.
Where AI could help
High AI leverage. The core is computation and data modeling, where ML demonstrably helps; synthesis prediction is weak and some novelty claims were overstated.
- Universal interatomic potentials that approach DFT accuracy at a tiny fraction of the cost
- Generative models that propose crystal structures for target properties
- Learn from failed syntheses and lab logs to predict recipes and kinetics
- Drive autonomous labs that test predictions at the same pace
Shown so far
- In November 2023 DeepMind reported that GNoME predicted 2.2 million new crystals, about 380,000 of them most stable, and said outside labs had independently made 736 of them. source
- In January 2025 Microsoft published MatterGen in Nature, a generative model that designs crystals for target properties; one synthesized material came within 20% of its target value. source
- The Matbench Discovery benchmark (August 2023) found universal interatomic potentials to be the best ML method for pre-screening stable crystals, with up to 6x faster discovery than random picking. source
Prerequisites
- Quantum Mechanics1925
- Quantum Chemistry & Bonding1927
- Density Functional Theory1965Predicting materials from first principles rests on DFT
- AI Discovers New Materials2023
Unlocks
- Materials Made to Order2040s?