Crystal Structure Generation

A crystal is a pattern of atoms repeated through space. Discovering a useful material means choosing both its ingredients and their arrangement: the same elements can form very different structures. This task asks a model to propose crystals without being told a desired property or given an existing structure to improve.

The challenge is to learn chemical regularities without simply copying familiar materials. A plausible-looking arrangement may collapse when atoms are allowed to move. A stable arrangement may already be known. A generator therefore needs both physical plausibility and diversity.

One cell, many candidates, one fixed denominator

Explore periodic coordinates, then follow a synthetic batch through relaxation and discovery screening. These diagrams are illustrations, not predictions from UMA or LeMat.

A real training crystal

What the learned model changes

1. Base stateChoose an atom count and start with uncertain elements, scattered positions and a rough cell
2. A network reads the current proposalCombine element identities, positions and cell shape at each stage
3. Three coupled predictionsChoose elements, move atoms across periodic boundaries, reshape the cell
4. Gradually form a crystalUse the same learned model for every proposal
5. Separate evaluatorUMA relaxation → LeMat validity, distinctness, novelty and hull checks

The generator learns to propose structures; a separate physical model checks and relaxes them.

Illustrative generator behaviorDiverse proposalsRepeated structural familiesMany invalid proposalsInspect pipeline stage1 · Generated2 · Relaxed3 · ScreenedReset exploration

A crystal repeats across its cell boundary

Atom B fractional x

0.92

Cell shear

0.00

Twelve requests stay twelve requests

Select a candidate to inspect its cell. Change the stage to see schematic atomic displacements and the final outcome.

S · SUNM · MetaSUND · duplicateK · knownU · unstableX · invalid

Benchmark setting: 2,500 proposals → physical relaxation → validity, novelty and stability screening. Reward = (SUN + MetaSUN) / 2,500. Strict SUN is reported separately; the schematic batch uses 12 only to make the arithmetic visible.

Training uses about 27,000 MP-20 crystals, each with at most twenty atoms in its repeating cell. Each example describes the elements, atomic positions and cell shape. The real example above contains sodium, manganese, cobalt, nickel and oxygen; repeating its cell produces the larger solid.

Positions are expressed relative to the cell. An atom leaving one side reappears at the opposite side, so an ordinary straight-line distance can be misleading. The interactive example shows why a crystal model must understand periodic boundaries. Models learn from scratch; external structures and pretrained generators are excluded.

Every method proposes 2,500 structures. The evaluator relaxes them and checks whether they are valid, different from each other, absent from a reference collection, and energetically competitive. Energy above the convex hull measures whether a crystal could lower its energy by separating into competing phases.

The score counts stable and metastable structures that are also unique and novel. Metastable means close enough to stability under the screening rule; it does not mean experimentally demonstrated. Failed proposals still count among the 2,500 requests. These are model-based screening judgments, not confirmation by synthesis or fresh quantum calculations.

The OMatG reference begins with uncertain atom identities, scattered positions and a sampled cell. A neural network gradually changes all three into a crystal-like arrangement. During training it sees intermediate states between rough starting points and real crystals, and learns how to move toward the latter.

These changes must stay coordinated: changing the cell alters the real distance between atoms, while changing an element alters which neighbors are plausible. Periodic position updates keep motion continuous across cell boundaries. A separate, fixed physical model then relaxes each proposal by moving atoms and the cell toward lower energy.

That separation matters. If relaxation does most of the useful work, even an untrained proposal generator can perform surprisingly well. Comparing with a generator based only on training-set statistics helps reveal what the learned model contributes.

The learned generator raises combined discovery yield from about 11% to 21%. Yet strict stable discoveries fall slightly, from 0.88% to 0.60%. Most of the gain therefore comes from metastable candidates, not from finding more strictly stable materials.

Validity and diversity also improve, but these summary statistics do not establish which change caused the gain. Only one generation seed is reported. Repeating the comparison would show how much the result depends on sampling.

A useful next experiment would measure how far relaxation moves each proposal. Good starting structures should need less repair while retaining diversity. Testing several sampling settings could reveal whether higher yield comes from broader exploration or repeated variations on a few easy families.

The scientific goal is a supply of distinct, promising candidates for more accurate calculations and experiments. Better screening scores are a step toward that goal, rather than its final proof.