CRUCIBLE: Methods for Quantum-Accelerated Materials Discovery

1Introduction and scope

The space of possible inorganic materials is astronomically large, and most of it has never been computed, let alone made. CRUCIBLE narrows that space in silico so that scarce experimental effort is spent only on materials worth synthesizing.

Combining the periodic table into stoichiometries and crystal structures yields a design space far beyond exhaustive enumeration, and each candidate traditionally costs minutes to days of density-functional theory (DFT) to evaluate. The job of computation is not to replace the furnace but to raise the prior probability that a synthesized material is stable, has the target property, and can actually be made. CRUCIBLE targets four decision points where that prior can be raised the most.

1.1Design philosophy

1. Learn the cheap, compute the fundamental. Structure-property relationships are learned from data; ground-state energetics of correlated systems, where DFT errs most, are computed with quantum chemistry.
2. Scientists in the loop. Every stage exposes an interpretable output and a gate at which a materials scientist decides. CRUCIBLE proposes; the lab disposes.
3. Uncertainty everywhere. No prediction ships without a calibrated confidence, so a low-confidence surrogate estimate is escalated to simulation rather than trusted.
4. Synthesis is ground truth. Models are judged against measurement, and every measurement is fed back to improve them.

1.2Scope and non-goals

This document covers crystalline inorganic materials, alloys, oxides, semiconductors, and related solid-state chemistries, against a target property profile. It does not cover molecular or biological systems (see Ace Hacker's PROTEUS engine), and it does not claim quantum hardware today outperforms classical methods on every task, §10 is explicit about where it does not. CRUCIBLE is a discovery accelerator, not an autonomous laboratory.

2Pipeline overview

The four methods form a loop. A target profile enters; a ranked shortlist of synthesizable candidates leaves; measured results retrain the learned components.

Sections 3–6 describe each method; Section 7 the active-learning loop; Sections 8–9 data and validation; Section 10 a candid account of limitations.

3Predict — property surrogates

Predict learns a surrogate for density-functional theory that reproduces its properties in milliseconds, so an entire composition space can be triaged before any expensive calculation runs.

3.1Crystal representation

A crystal is periodic, so it is encoded not as a finite molecule but as a graph with periodic boundary conditions: nodes are atoms in the unit cell, edges connect an atom to its neighbors including their periodic images within a cutoff radius, and edge features encode interatomic distances and directions. This representation respects the translational symmetry that defines a solid.

3.2Message-passing model

A graph neural network refines each atom's representation by aggregating messages from its neighborhood; after K rounds, a pooled crystal embedding predicts the target property. One update step is

where M is a learned message function and e_ij the edge features. Distance-based, rotation-invariant edge features make the predictor invariant to how the cell is oriented, a physical requirement, not a nicety.

3.3Multi-property and uncertainty

One model predicts formation energy, band gap, bulk and shear moduli, and thermal conductivity jointly, sharing a learned representation so data-rich targets support data-poor ones. Predictions carry calibrated uncertainty from a deep ensemble; a candidate whose uncertainty exceeds a threshold is routed to quantum simulation (§4) rather than trusted on the surrogate alone.

4Simulate — quantum electronic structure

Where a surrogate is uncertain and where classical DFT is unreliable, correlated electrons, magnetism, near-degenerate states, CRUCIBLE computes electronic structure with a variational quantum eigensolver on the decisive orbitals.

4.1Active-space embedding

A full solid is far too large for a quantum device. CRUCIBLE isolates the strongly correlated active space, the orbitals near the Fermi level that govern the property of interest, and treats it quantum-mechanically while embedding it in the classically-treated environment via density-matrix embedding theory (DMET). Only the active space's electronic Hamiltonian is sent to the quantum solver.

4.2Variational quantum eigensolver

The active-space electronic Hamiltonian, in second-quantized form, is

mapped to qubits (Jordan–Wigner or Bravyi–Kitaev). VQE prepares a parameterized trial state and uses the variational principle, the energy expectation is an upper bound on the true ground state, so minimizing it approaches the answer:

The trial state uses a unitary coupled-cluster (UCCSD) ansatz built on the mean-field reference, |ψ(θ)⟩ = e^{T(θ)−T†(θ)}|Φ₀⟩. A classical optimizer proposes parameters; the quantum backend estimates the energy by measurement; the loop repeats to convergence.

4.3From energies to properties

Band gaps follow from the difference of ground-state energies of the neutral and charged cells; magnetic order from the energies of competing spin configurations. Phase stability follows from formation energy. The per-atom formation energy of a compound relative to its elemental references μ_i is

4.4Error mitigation and validation

On real hardware CRUCIBLE applies readout-error calibration and zero-noise extrapolation, and validates against high-accuracy classical references (coupled cluster, and where tractable full configuration interaction) and experimental data. Where quantum execution is not yet advantageous, the same interface runs a correlated classical method and records which was used.

5Design — inverse design

Screening a database only finds what is already in it. Design runs the problem backwards: given a target property profile, a generative model proposes novel crystals that should meet it.

5.1Symmetry-aware crystal diffusion

Crystals are generated by a diffusion model that jointly produces atom types, fractional coordinates, and the lattice, respecting the periodic and point-group symmetries that make a structure physical. Training corrupts a known structure with noise; generation reverses it. The forward step is the standard

and a network ε_θ, made equivariant to rotations, translations, and periodicity, learns to denoise, so generated structures are symmetry-consistent by construction rather than by post-hoc filtering.

5.2Property-conditioned, multi-objective

Generation is conditioned on the target profile and steered toward a weighted objective over interpretable properties, evaluated with the fast surrogate of §3:

Weights are set per program, a transparent conductor and a hard structural alloy want different profiles. Candidates are pushed toward unexplored chemistry (novelty) while staying near the convex hull (stability), invention that can actually exist.

Generation loop (simplified)for target in profiles:
    cands = diffusion.sample(target, n=20_000)     # symmetry-aware
    cands = [m for m in cands if valid(m) and novel(m)]
    scored = [(m, reward(m)) for m in cands]      # surrogate-evaluated
    top = select_pareto(scored, k=2000)             # multi-objective
    return [m for m in top if near_hull(m)]        # plausibly stable

6Screen — stability and synthesizability

A generated candidate is worthless if it decomposes or cannot be made. Screen ranks the field by thermodynamic stability, synthesizability, and cost, cutting a vast space to the few materials a lab should attempt.

6.1Thermodynamic stability

A material is stable only if no combination of competing phases at the same composition has lower energy. Its distance above the convex hull of formation energies quantifies this; a material on the hull (E_hull = 0) is thermodynamically stable, and a small positive value indicates a possibly-synthesizable metastable phase:

where Ê_hull(c) is the hull energy interpolated at composition c. Hull distance is the first, cheapest filter, computed from surrogate or simulated formation energies.

6.2Synthesizability and cost

• Synthesizability — a learned model estimates whether a plausible route from available precursors exists, trained on the historical record of what has and has not been made.
• Earth-abundance and toxicity — candidates are weighted by element scarcity and hazard, so a discovery is also a viable material, not a lab curiosity built on a rare element.
• Applicability domain — a candidate far from the training distribution is flagged as an extrapolation rather than scored with false confidence.

6.3Calibrated triage

Filters run cheapest-first, surrogate stability, then synthesizability, then quantum verification on survivors, so expensive computation only touches promising candidates. Every score carries a conformal uncertainty; for target coverage 1−α, the prediction set satisfies

so a candidate whose stability interval straddles the decision threshold is escalated to simulation or to a human, never silently passed.

7The active-learning loop

CRUCIBLE is not a one-shot predictor. Each campaign produces synthesis and characterization data, the most valuable signal in the system, and the loop chooses what to make next to learn the most.

7.1Closing the loop

Synthesized materials are characterized for structure, stability, and the target property. Results are written back to the training corpus with full provenance, and Predict, Design, and Screen are periodically retrained. Simulate, being first-principles, is not retrained but is recalibrated against measured properties.

7.2Acquisition

Which candidates to synthesize next is a design-of-experiments question. CRUCIBLE selects a batch balancing exploitation (high predicted value) against exploration (high model uncertainty), maximizing an acquisition over the candidate pool:

where μ is predicted value, σ its uncertainty, κ trades the two, and a redundancy penalty keeps the batch chemically diverse so a synthesis round is not spent on near-duplicates.

8Data and reproducibility

Method quality is bounded by data quality and by the discipline of the splits used to measure it. CRUCIBLE is explicit about both.

8.1Training data

8.2Splits and leakage

Random splits flatter models by leaking near-neighbors between train and test. CRUCIBLE reports on the splits that matter for real deployment:

• Composition split so the test set contains genuinely different chemistries, not stoichiometric neighbors of training points.
• Structure-prototype split holding out entire structural families to test generalization beyond seen prototypes.
• Temporal split training on materials known before a cutoff and testing on those characterized after, mirroring prospective discovery.

8.3Reproducibility

Data snapshots, DFT settings, model weights, and configs are versioned together, and every reported number is tied to a specific pipeline version and split so it can be reproduced.

9Validation and benchmarks

The only validation that counts is prospective: does a material the model liked behave as predicted when it is synthesized and measured? We report retrospective benchmarks to calibrate expectations and prospective results to prove them.

9.1Metrics

Property prediction is scored by MAE against DFT and experiment; stability classification by the precision of "predicted stable" against measured hull distance; generative design by validity, novelty, and the hit rate of stable, on-target candidates; simulation by MAE against high-accuracy references.

The verdict column is the point: the learned methods carry most of the retrospective lift now; quantum helps on the correlated simulation step in simulation and on small active spaces today.

10Limitations and current maturity

This section is deliberately blunt. Computational materials discovery has overpromised before; credibility with experimentalists depends on candor.

• NISQ reality. Quantum hardware is noisy and small. Simulate's advantage is established on modest active spaces and in simulation; larger systems rely on classical correlated methods, and CRUCIBLE routes to them automatically.
• DFT reference error. Surrogates trained on DFT inherit its systematic errors; where DFT itself is wrong, the surrogate is confidently wrong. This is precisely why correlated cases are escalated to Simulate and, ultimately, to the lab.
• Stability is necessary, not sufficient. A material on the convex hull can still resist synthesis for kinetic reasons the thermodynamics do not capture; the synthesizability model is a prior, not a guarantee.
• Generative models hallucinate. Novelty and near-hull filters plus human review are mandatory because generative models can propose confident nonsense.
• Simulated figures. Every number here is illustrative of the target. Production performance depends on chemistry, data, and backend, and is established per program and validated by synthesis.

11Responsible use and intellectual property

Materials design is dual-use. CRUCIBLE is operated for civilian research under partner agreements, with controls appropriate to the chemistry involved and safeguards against the design of hazardous materials. Partner data and generated candidates remain the partner's intellectual property; deployment is in a secure enclave, and models trained on partner data are not shared across partners. Provenance is retained so inventorship and the origin of every proposed material can be established.

12Roadmap

—References and further reading

1. Xie & Grossman. Crystal Graph Convolutional Neural Networks. Phys. Rev. Lett. 120 (2018).
2. Chen et al. Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals (MEGNet). Chem. Mater. 31 (2019).
3. Jain et al. The Materials Project. APL Materials 1 (2013).
4. Xie et al. Crystal Diffusion Variational Autoencoder for Periodic Material Generation (CDVAE). ICLR 2022.
5. Peruzzo et al. A variational eigenvalue solver on a photonic quantum processor. Nature Communications 5 (2014).
6. Knizia & Chan. Density Matrix Embedding Theory. Phys. Rev. Lett. 109 (2012).
7. Bartel et al. A critical examination of compound stability predictions. npj Comput. Mater. 5 (2019).
8. Angelopoulos & Bates. A Gentle Introduction to Conformal Prediction. 2021.
9. Lan et al. AdsorbML / Open Catalyst and high-throughput screening at scale. 2023.
10. Merchant et al. Scaling deep learning for materials discovery (GNoME). Nature 624 (2023).

This document describes methods and design intent. All performance figures are illustrative and simulated for demonstration; production characteristics are established per program and validated experimentally. © Ace Hacker Research & Development Lab.