From Crystal to Candidate: Materials Science on Neurosnap

A crystal structure is a starting point. The interesting questions begin when you ask how it behaves.

A promising composition can hide a difficult material. A lattice may relax into a different geometry. An apparently rigid solid may have a soft vibrational mode. A vacancy can open a migration pathway, while a new surface exposes an entirely different arrangement of atoms.

For researchers working on batteries, catalysts, semiconductors, and functional solids, these details shape the next experiment. Which candidate deserves a closer look? What should be measured? Which assumption needs checking before a larger calculation?

Neurosnap is expanding into materials science with eleven tools for periodic crystal structures: preparation, generation, relaxation, dynamics, elastic response, phonons, phase stability, defects, surfaces, diffraction, and migration barriers. Together, they make it easier to move from a structural candidate to a more specific scientific hypothesis.

Figure 1. A question-first guide to the materials collection. These are complementary tools you can combine as appropriate; the diagram does not represent an automatic screening pipeline or a single measure of stability.

Give the structure a proper introduction

Before comparing properties, it helps to know what structure you actually have.

Crystal Preparation & Analysis inspects symmetry, compares structures, standardizes primitive or conventional cells, and builds supercells. This is useful when two files look different but may describe the same periodic arrangement—or when a calculation needs a larger cell than the one you downloaded.

Crystal inputs now have their own place on Neurosnap. Upload an ordered CIF or VASP 5 POSCAR, inspect composition and lattice information, and open a 3D preview. Single-cell and supercell views help reveal the periodic arrangement before you submit a calculation. Basic checks catch common file and geometry problems; each calculation applies its own further validation and size limits.

Next comes Crystal Relaxation: optimize atomic positions with the lattice fixed, or relax positions and the cell together. Choose a supported machine-learned interatomic potential or a PBE DFT method, then examine energy, force, stress, and convergence information alongside the final geometry.

That distinction matters. A relaxed structure is a local minimum candidate under the selected model. It does not, by itself, establish experimental stability, synthesizability, or resistance to decomposition.

If you are looking for candidates rather than starting from an existing structure, Crystal Generation offers two routes. MatterGen proposes structures conditioned on a chemical system or a target band gap; symmetry-based generation uses an integer composition, space group, and volume estimate. The underlying MatterGen research provides the generative foundation. A conditioning target guides sampling—it is not a verified property of the returned candidate. Generated structures still need relaxation and independent evaluation.

Ask how the crystal responds

A structure becomes more informative when you perturb it.

Crystal Elasticity & Equation of State explores small strains or changes in volume. It returns elastic tensors and moduli, or an energy–volume curve with a fitted bulk modulus. These can help compare candidate stiffness or compressibility. The elastic-tensor calculation uses a clamped-ion finite-strain treatment at the relaxed cell; the equation-of-state workflow relaxes internal coordinates at the sampled volumes. Inspect fit quality and sensitivity to the chosen strain range before treating a small difference as meaningful.

Crystal Phonons & Thermal Properties probes collective atomic motion with finite-displacement forces and Phonopy. Phonon bands, a vibrational density of states, and harmonic thermal properties provide another perspective on the same crystal. Imaginary frequencies can flag a problem with the assumed structure or calculation settings; supercell and reciprocal-space convergence remain essential. The workflow withholds thermal results when its integration mesh contains sufficiently imaginary modes.

Powder XRD Simulation connects the atomic arrangement to a familiar experimental observable: a predicted powder diffraction pattern. It is a useful way to explore how crystal geometry maps onto peak positions and relative intensities. The displayed broadening is illustrative; it does not infer crystallite size, instrument response, or experimental phase fractions.

Figure 2. Three views of silicon, drawn from actual saved service outputs. Left: the supplied 5.43 Å cubic diamond-silicon structure gives a Cu Kα (111) peak at approximately 28.47° in 2θ; the curve uses the service's illustrative Gaussian broadening. Center: seven MatterSim energy–volume samples after starting-cell relaxation, plotted relative to their lowest sampled energy; the reported Birch–Murnaghan fit gives a bulk modulus of approximately 85.85 GPa. The line connects samples and is not the fitted equation of state. Right: MatterSim/Phonopy phonon branches along Γ–X from a small demonstration calculation using a 1 × 1 × 1 primitive-cell supercell and a 3 × 3 × 3 integration mesh. These are calculation examples, not experimental comparisons or converged reference values; each workflow prepares its geometry independently.

One material, several questions. A recognizable diffraction pattern tells you something different from a good energy–volume fit, and neither replaces a vibrational or thermodynamic assessment.

Follow atoms, not just snapshots

A static structure cannot tell the whole story of atomic motion.

Materials Molecular Dynamics uses MatterSim forces to simulate fixed-cell NVT or NVE dynamics. Trajectories, energies, radial distribution functions, and mean-squared displacement help you examine what happens over the simulated time window. The MatterSim paper describes the underlying learned atomistic model; its predictions still need scrutiny for your chemistry and conditions.

For transport studies, the temptation is to turn every rising displacement curve into a diffusion coefficient. Short trajectories can be misleading. Neurosnap withholds a coefficient unless the trajectory passes minimum sampling and diffusive-regime checks. Even an accepted estimate remains subject to finite-cell, timescale, and potential-model errors.

Crystal Migration Barriers asks a more focused question: what modeled energy barrier separates two specified, atom-mapped crystal configurations? A climbing-image nudged elastic band calculation estimates a minimum-energy path between compatible endpoints. This is useful for investigating a proposed vacancy hop or other localized rearrangement. Endpoints must have matching cells, species, and atom order, and should already be relaxed using the same potential. A potential-energy barrier is not a measured rate or a full activation free energy; check the returned convergence information before interpreting it.

Change one feature and see what question appears

Sometimes the most interesting material is one atom away from the structure you started with.

Crystal Defect Builder creates vacancies, substitutions, and specified interstitials. Surface & Slab Builder generates periodic slab terminations for selected Miller indices, with controllable thickness and vacuum separation.

These tools make structural hypotheses tangible. You can inspect the vacancy, identify which atomic planes a surface exposes, and take the resulting geometry into a suitable next calculation. Building a defect does not calculate its formation energy or charge state. Building a slab does not establish its reconstruction, surface energy, polarity, or catalytic activity.

Figure 3. A structural hypothesis you can inspect. Left: the eight-atom conventional silicon cell. Center: the corresponding vacancy example contains seven atoms; the outlined marker identifies the removed site's original position and is not an atom in the output. Right: the generated (100) slab is shown as stored, with its periodic cell and vacuum region. The vacancy and slab are unrelaxed construction examples. Atom sizes are illustrative; these views do not imply defect energetics or surface stability.

For a different kind of comparison, Phase Stability & Decomposition constructs a convex hull from the phase energies you supply. It reports energy above that supplied hull and possible decomposition products. This helps organize competing-phase evidence, provided the energies are comparable and the relevant elemental references and competing phases are included.

The word supplied is doing real work here. Missing a competing phase can change the conclusion. Energies from different methods or inconsistent reference conventions should not be mixed into a single stability claim.

How the pieces fit together

A practical study might begin with a CIF, standardize the cell, relax it, and then branch according to the question. A transport project might build a defect and investigate a migration path. A mechanical study might compare elastic response across relaxed candidates. A structural assignment might start with a powder diffraction simulation. Generated candidates can enter the same process after appropriate preparation and relaxation.

At a high level, the approach is consistent: validate the periodic structure, prepare the geometry needed by the chosen method, perform the calculation, and inspect numerical results alongside the structure. Periodicity, atom mapping, convergence, units, and reference choices matter throughout. A successful job is the beginning of interpretation.

The collection builds on established materials software, including pymatgen, symmetry analysis with spglib, atomistic calculations with ASE, and workflow-specific methods such as Phonopy and MatCalc. Learned potentials provide an efficient route to forces and energies for several workflows; Crystal Relaxation also offers selected DFT methods. These methods answer different questions and should be compared with their assumptions in view.

What have we checked?

We exercised the workflows with real crystal inputs and inspected the structures and numerical outputs they returned. For conventional diamond silicon, preparation recovered the expected eight-atom cell and Fd-3m symmetry. The diffraction example reproduced the expected geometric location of the first Cu Kα peak. Defect construction changed the atom count as intended, and slab generation produced readable periodic structures with positive surface areas.

Further checks covered energy–volume and elastic outputs, phonon and thermal results, NVT and NVE trajectories, sampling intervals that do not divide the requested run length, converged migration-path examples, and both symmetry-based and MatterGen generation. We checked rejection of unsupported or disordered structures, invalid geometries, and out-of-range settings. Saved calculation outputs also exercised the result tables, plots, and structure displays.

Those checks give us evidence that the workflows carry the intended scientific quantities through a calculation. They do not establish predictive accuracy across arbitrary materials. The silicon examples here are demonstrations, and the phonon example deliberately uses small sampling settings. Before relying on a result, compare against appropriate measurements or trusted reference calculations, test convergence, and check whether the model and assumptions fit your system.

Bring a sharper question to the next material

Which candidate should you relax? Which structural difference deserves a measurement? Does a proposed defect create a plausible migration path? Is a reported stability result supported by the competing phases you included?

The new collection gives scientists more ways to ask those questions—and more concrete outputs to take into the next stage of research.

Explore the Neurosnap materials science tools, and revisit our quantum chemistry collection for molecular electrostatics, solvent effects, absorption, and vibrations. Our formulation overview connects those molecular questions to solubility, ionization, and mixture thermodynamics.

Figures are original Neurosnap illustrations. Figures 2 and 3 use saved calculation outputs; Figure 1 is a conceptual guide. No experimental benchmark data are presented.