Getting Started

Spectra is an AI-driven materials discovery platform. Describe the material you need in plain language and the system will route your request to the best generative models, rank candidates by structural quality, and provide deep computational analysis.

1. Describe your target material

Navigate to the Search page and enter a natural-language description of the material you need. For example:

"High bulk modulus ceramic for thermal barrier coating"
"Novel lithium-ion battery cathode with low formation energy"
"Magnetic intermetallic with high Curie temperature"

You can also click a category chip on the landing page to pre-select a domain (e.g. Superconductors, Battery Materials).

2. Review generated candidates

Spectra routes your request to the optimal model variants, generates candidate structures, and ranks them by a confidence-weighted score. Results appear in the Results page with interactive 3D crystal viewers and per-candidate property cards.

3. Run computational analyses

Use the Analyze ▾ dropdown on any candidate to run structural analyses:

  • Symmetry — space group, crystal system, Wyckoff sites
  • Voronoi — coordination environments per atom
  • Bond Distribution — per-pair bond statistics and histograms
  • DFT Surrogate — fast CHGNet prediction of formation energy, stability, forces, stress, magnetic character
  • Phase Stability — convex hull analysis against known Materials Project phases
  • Real DFT — full GPAW calculations: single point, relaxation, elastic constants, DOS, band structure

4. Generate reports

Use the Reports ▾ dropdown to generate professional documents:

  • Scientific Document — comprehensive structural and property summary
  • Economic Report — cost analysis, supply risk, recyclability, market positioning
  • Patent Packet — patent-grade writeup with prior art screening (PubChem + USPTO)

5. Save and compare

Click the ★ button on any candidate to save it. Saved candidates persist across sessions and appear on the Saved page. Use research notes to annotate candidates with observations for your team.

Supported models

Spectra currently integrates the following generative models. See Model Benchmarks for published performance metrics.

  • MatterGen — diffusion-based inorganic crystal generation (Microsoft, Nature 2025)
  • IBM FM4M — foundation model for molecules: SMI-TED, BART, MHG variants. IBM archived the upstream repository on 2026-10-07; the published weights Spectra runs remain available under Apache-2.0, so these capabilities are unaffected.
  • GT4SD: property-conditioned molecular generation via the Regression Transformer. Give it a target logP from -5 to 10 and it edits a seed molecule toward it; every molecule comes back with its measured logP and the signed distance from your target. Experimental: targets above about 3.4 logP are accepted but rarely reached from a seed, and IBM's own model card lists production-level inference as out of scope. See Model Benchmarks for per-variant availability.