Live discovery console • simulated data

Most of the materials that could exist don’t yet.

CRUCIBLE is Ace Hacker's materials-discovery engine. It predicts a material's properties before it is made, simulates its electrons with quantum chemistry, designs new compositions to hit your targets, and screens for what is stable and synthesizable, turning materials discovery from search into design.

Request research access → Explore the engine

0.04 eVper-atom formation MAE

10⁴×faster than DFT

340Mcompositions screened

crucible://simulate/unit-cell RELAXED

crystal unit cell

formation energy−0.42 eV/atom

Composition

Fe₂VAl

Band gap

1.1eV

Hull distance

0meV

Engine at a glance

Candidate compositions screened in silico

Screen

0 eV

Formation-energy MAE vs DFT (per atom)

Predict

Novel stable materials on the convex hull

Simulate

Faster than first-principles per candidate

End to end

One engine, four instruments

The whole discovery loop, from target to crystal.

CRUCIBLE pairs each step with the right computation, deep learning where a surrogate beats brute-force simulation, quantum chemistry where electrons must be treated honestly, and keeps a scientist in the loop at every gate.

Predict

Property surrogates

ML models predict formation energy, band gap, elastic and thermal properties in milliseconds, not CPU-days.

graph networks • DFT-trained

Simulate

Quantum structure

VQE computes electronic structure and phase stability where classical DFT approximations break down.

VQE • band structure • hull

Design

Inverse design

Generative models invent new compositions and structures shaped to your target property profile.

diffusion • multi-objective

Screen

Stability & synthesis

High-throughput screening ranks candidates by stability, synthesizability, and cost before any furnace runs.

convex hull • synthesizability

Predict • Instrument 01

A property in milliseconds, not CPU-days.

Density-functional theory is accurate but slow, minutes to days per material. CRUCIBLE learns a surrogate from millions of DFT calculations that reproduces the answer in milliseconds, so an entire composition space can be evaluated before a single furnace is lit.

01
Structure-aware graph networksCrystals are encoded as periodic graphs; message passing over atoms and bonds captures the local chemistry that sets properties.
02
Multi-property, one modelFormation energy, band gap, bulk modulus, and thermal conductivity are predicted jointly, sharing a learned representation.
03
Calibrated uncertaintyEvery prediction carries an error bar, so a low-confidence estimate is sent to simulation rather than trusted blindly.

Crystal graph netsDFT-trainedEnsemble UQMaterials Project

Predicted vs DFT • formation energyR² = 0.994

predicted (eV/atom) vs reference

Held-out materialparity line

Prediction error distribution|pred − DFT|

Predicted properties • lead candidatenormalized

Simulate • Instrument 02

Where a surrogate isn’t enough, simulate the electrons.

Correlated electrons, magnetism, and near-degenerate states are exactly where classical DFT is least reliable, and where a material's most interesting behavior lives. CRUCIBLE computes the electronic structure of the decisive orbitals with a variational quantum eigensolver.

Electronic band structureE vs k • gap 1.1 eV

energy along high-symmetry path

Valence bandsConduction bandsBand gap

01
Active-space embeddingThe strongly correlated orbitals are treated on the quantum device; the rest is embedded classically, so problems fit today's hardware.
02
VQE ground statesA variational eigensolver finds ground-state energies, from which band gaps, magnetic order, and phase stability follow.
03
Phase stabilityFormation energies across a composition place each candidate on the convex hull, the line between stable and metastable.

VQE / UCCSDDMET embeddingerror mitigationphonons

VQE energy convergenceHartree

Convex hull • phase stabilityEf vs composition

stable = on hull

Phase diagram • temperature × compositionstable phase by region

Design • Instrument 03

Don’t screen the known. Invent the unknown.

Screening a database can only find what is already in it. CRUCIBLE runs the problem backwards: given a target property profile, a generative model proposes novel compositions and crystal structures that should meet it, including chemistries no one has catalogued.

Composition space • generated candidates6,400 structures

latent projection

GeneratedPareto-optimalSelected

01
Crystal diffusionA diffusion model generates atoms, lattice, and symmetry together, so structures are physically plausible by construction.
02
Property-conditionedGeneration is steered toward a target band gap, stiffness, or conductivity, and multiple objectives at once.
03
Novel yet stableCandidates are pushed toward unexplored chemistry while staying near the convex hull, invention that can actually be made.

Equivariant diffusionSymmetry-awarePareto searchactive learning

Target property matchcandidate vs target

Property gain over generationsactive-learning loop

Screen • Instrument 04

Millions of candidates. A handful worth making.

A generated candidate is worthless if it decomposes or can't be synthesized. Screen ranks the field by thermodynamic stability, synthesizability, earth-abundance, and cost, cutting a vast space down to the few materials a lab should actually attempt.

Element frequency in stable candidates • periodic tableoccurrence • hover to inspect

RareCommonFrequentDominant

01
Stability firstDistance to the convex hull filters out anything that would decompose into competing phases.
02
SynthesizabilityA learned model estimates whether a known route could plausibly make the material, and suggests precursors.
03
Abundance & costCandidates are weighted by element scarcity and toxicity, so a discovery is also a viable material.

Hull distanceSynthesizabilityEarth-abundancecost model

Discovery funnelper campaign

Property map • stiffness vs density (Ashby)by material class

Benchmarks • honest numbers

Where quantum earns its place, and where it does not.

We benchmark against strong classical baselines, tuned DFT and state-of-the-art ML potentials. The surrogate wins on throughput; quantum simulation wins on the correlated systems DFT gets wrong. CRUCIBLE routes each material to the cheapest method that is accurate enough.

Accuracy vs method • correlated oxidesMAE meV/atom • lower better

Baselinechemical-accuracy line

Capability profilevs classical HT-DFT stack

Architecture • from target to crystal

One loop, closed by the lab.

A target profile enters; a ranked shortlist of synthesizable candidates leaves. Measured results from the lab flow back to retrain the surrogates, so the engine sharpens with every campaign.

Now partnering on programs

Bring us a property target. We’ll bring the material.

CRUCIBLE is in research partnerships with a small number of teams in energy, semiconductors, and structural materials. If you have a hard property target, we should design against it together.

Request research access → Read the methods paper

Secure enclave deployment • your IP stays yours • wet-lab validation partners on request