from first research to first use: the typical timeline for a material.
Materials Genome Initiative white paper, 2011.
Trial-and-error synthesis, sparse data, and some 10¹⁰ chemically plausible four-element compositions to choose from (Davies et al., Chem 1, 617, 2016). We move the first pass of that search into computation.
Energy storage · Semiconductors · Coatings · Catalysis · Solar · Thermal management · Construction
Astralby Accretion Labs
Describe the material you need. Astral returns a ranked shortlist and tells you how far to trust each answer.
Boron arsenide, a 3 × 3 × 3 zinc blende supercell at its measured lattice constant: boron in pearl, arsenic in polished copper. An illustration, not an Astral result.
01
Ask
Write what you need in plain language; Astral turns it into criteria you can edit, and marks anything it assumed.
02
Build
It proposes candidate crystals derived from known structures.
03
Check
Open, pretrained AI models estimate stability and key properties.
04
Rank
Candidates are ranked against your goals, each with a short reason.
05
Explain
Every result carries a confidence label, and every number shows where it came from: AI prediction, quantum calculation or database.
0.035eV/atom
95% CI 0.025–0.049 · n = 497 of 500
Average gap between the stability predicted by an open, pretrained model that Astral runs and quantum (DFT) calculations, on 500 structures outside Materials Project, in chemical systems not used to tune Astral. 497 of the 500 could be scored.
15of 15
95% CI 80–100% · n = 15
Starting from 7 known compounds, Astral rebuilt and ranked all 15 compounds that Materials Project already lists as stable. These are predicted rankings, compared with database values.
Accretion Labs internal validation, October 2026. Predictions, not lab results.
Predictions, not lab results. Astral doesn't invent synthesis routes.
Data: the Materials Project, computed properties for 150,000+ inorganic materials, used under CC BY 4.0; its GNoME-derived subset (CC BY-NC) is excluded. Jain et al., APL Materials 1, 011002 (2013).
Screen a perovskite yourself.
This one runs in your browser. Pick a cation for each site and an anion; the page computes three descriptors from Shannon ionic radii (Kieslich effective radii for MA⁺ and FA⁺) and runs two tests: Goldschmidt's 1926 tolerance factor with the octahedral factor, and Bartel's 2019 τ. A worked example using a published rule (Shannon radii, Bartel et al. tolerance factor), not Astral's method
X · anion
A · large cation, 12-coordinate
B · small cation, 6-coordinate
SrTiO₃
Goldschmidt tolerance factor t
0.70perovskite if 0.825 < t < 1.0591.20
octahedral factor μ = rB / rX
0.20B fits an octahedron if μ > 0.4140.80
Bartel τ
1perovskite if τ < 4.189
Drag to rotate
ABXA copper · B core with disk · X pearl with electron cloud · ideal cubic cell, spacing from the radii, ions at 60% size
153 charge-balanced compositions are waiting (hybrid-cation oxides left out).
τ says perovskiteτ says notknown perovskite at room temperatureknown non-perovskitetap a point to load it
Radii: R. D. Shannon, Acta Cryst. A32, 751 (1976); methylammonium and formamidinium effective radii: G. Kieslich, S. Sun, A. K. Cheetham, Chem. Sci. 5, 4712 (2014). τ: C. J. Bartel et al., Sci. Adv. 5, eaav0693 (2019), who report 92% accuracy for τ against 74% for t on 576 compounds; the t window 0.825 to 1.059 is also from their fit. Both were fitted on inorganic cations only (largest A: Cs⁺, 1.88 Å), so for MA⁺ and FA⁺ the verdicts are an extrapolation. Fe³⁺ uses the high-spin radius. Known structures are a short hand-picked set from the literature, labelled by their thermodynamically stable room-temperature structure (so CsPbI₃ and RbPbI₃ count as non-perovskites), not a benchmark.
Preprints from the lab.
Sole-authored and not yet peer reviewed. Linked here because claims should be checkable; the code for both is public so anyone can rerun them.
preprint · SSRN · 2026
What Does It Cost to Validate a Predicted Inorganic Material?
An end-to-end cost model from a computational prediction to a sample made and checked in the lab: compute, synthesis attempts, characterisation, and the failure rates at each stage. The number that decides whether AI-for-materials is a business or a demo.
arXiv:2610.33101 · cond-mat.mtrl-sci · 22 Oct 2026
Chemistry-resolved disagreement between Materials Project-compatible PBE(+U) and r²SCAN stability verdicts in the GNoME materials database.
GNoME's stability labels rest on a GGA/GGA+U workflow, and its authors report that 84% of the binaries and ternaries it predicted stable retain a negative decomposition energy under r²SCAN. Across the 139,541 GNoME materials released with r²SCAN recomputations, the audit resolves where the two verdicts disagree by chemistry: among quaternaries alone, from 4.0% for plain intermetallics (no lanthanide, no 3d metal) to 30.7% for lanthanide-free oxides containing 3d metals, and disagreement is strongly associated with Materials Project Hubbard-U elements. Code and the full reproduction pipeline are public.
Others' published work that this field, and Astral, builds on.
DOE · Lawrence Berkeley National Lab
2011
The Materials Project
A Department of Energy–funded open database of computed properties for 150,000+ inorganic materials, run on DOE supercomputers. The public data foundation most later generative and graph-network models are trained or benchmarked against.
MIT
2018
CGCNN
Crystal graph convolutional neural networks (Xie and Grossman, Phys. Rev. Lett. 120, 145301, 2018) learn a crystal's properties directly from how its atoms are connected, trained on Materials Project calculations.
NIST
2021
ALIGNN
A graph neural network that adds bond angles to the graph, trained on NIST's JARVIS-DFT, the Materials Project and QM9 to predict 52 solid-state and molecular properties.
Google DeepMind
2023
GNoME
A graph-network model for materials exploration expanded the catalogue of stable inorganic crystals by roughly an order of magnitude: 381,000 entries added, about 421,000 predicted stable in total (Nature 624, 80, 2023).
Microsoft Research
2024
MatterSim & MatterGen
MatterSim, released in 2024, is a machine-learned force field that simulates materials across temperatures and pressures. MatterGen (preprint December 2023, Nature 2025) is a diffusion model: state a target property and it proposes an atomic arrangement designed to hit it.
Cambridge · Berkeley
2025
Matbench Discovery
An open benchmark that tests whether machine-learning models can tell stable crystals from unstable ones before any DFT is run (Riebesell et al., Nature Machine Intelligence, 2025), with a public leaderboard.
381,000stable crystals added by GNoME's predictions
Known from experiment (ICSD) and computationally stable20,000
Stable by computation, before GNoME48,000
Predicted stable after GNoME, in total421,000
Stable inorganic crystals. Sources: Merchant et al., Nature 624, 80 (2023); Google DeepMind, “Millions of new materials discovered with deep learning”, 29 November 2023. The GNoME figures are predictions; our audit maps where GNoME's PBE and r²SCAN verdicts disagree.
Bring us a materials problem worth solving.
A property target, a formulation problem, a dataset you want modelled. Two sentences is enough to start.