Model Benchmarks

Published prediction accuracy and generation quality metrics for every model and checkpoint variant used in the Spectra platform. All numbers are published results, sourced from peer-reviewed papers or official repositories, not measurements of a Spectra run. Source links are provided on each page, and each variant is labelled with whether Spectra can currently run it.

MatterGen

Crystal Generation

Nature (2025) →GitHub →

Diffusion-based inorganic crystal generator conditioned on target properties. Trained on 608,000 stable materials from Materials Project + Alexandria databases.

MatterGen Base (Alex-MP)

Unconditional generation trained on the full Alex-MP dataset (608k stable materials).

Structure Validity: 100%Composition Validity: 88.5%Uniqueness: 99.9%

MatterGen (MP-20)

Base checkpoint trained on MP-20 subset (45,231 structures, ≤20 atoms).

Structure Validity: 100%Composition Validity: 84.5%Uniqueness: 99.8%

Band Gap Conditioning (DFT)

Fine-tuned to generate crystals matching a target DFT band gap value.

Structure Validity: 100%

Bulk Modulus Conditioning (ML)

Fine-tuned to generate crystals matching a target ML-predicted bulk modulus.

Structure Validity: 100%Experimental Validation: TaCr₂O₆

Chemical System Conditioning

Generates crystals constrained to specified element sets.

S.U.N. (well-explored): 83%S.U.N. (partially explored): 65%S.U.N. (unexplored): 49%

Composition + Stability

Chemical system conditioning with energy-above-hull stability target.

Magnetic Density Conditioning

Generates crystals conditioned on target magnetic density (μB/ų).

Validation: 18 S.U.N.

Space Group Conditioning

Generates crystals constrained to a target space group (1–230).

Coverage: All 7 lattice typesSG Tested: 14 space groups

IBM FM4M (SMI-TED)

Molecular Generation & Property Prediction

arXiv (2024) →GitHub →

IBM Foundation Model for Materials — pre-trained on 91M SMILES from PubChem. BART, MHG, and SMI-TED model variants for molecular generation and featurization. IBM archived the upstream FM4M repository on 2026-10-07, making it read-only with no successor announced; the published model weights remain available under Apache-2.0 and are what Spectra runs, so these capabilities are unaffected.

SMI-TED 289M

Large encoder-decoder foundation model (289M params) pre-trained on 91M SMILES. Generation + property prediction.

Validity (fragment): 99.99%Scaffold Novelty: 99.99%Internal Diversity: 85.65%

BART Generative

De-novo molecule generation via BART autoregressive model.

MHG Generative

Molecular Hypergraph Grammar — structure-aware generation preserving ring systems.

GT4SD

Property-Conditioned Molecular Generation

npj Computational Materials (2023) →GitHub →

IBM Generative Toolkit for Scientific Discovery — wraps multiple algorithms with harmonized interfaces. Experimental in Spectra: property-conditioned generation through GT4SD is not yet verified in production, and the toolkit's GuacaMol, REINVENT and MoLeR algorithms are not wired into Spectra. The benchmark numbers below are the upstream published results, not measurements of a Spectra run.

Regression Transformer

Experimental

Property-conditioned generation — specify a target logP and the model edits a seed molecule toward it, jointly predicting and generating.

QED Prediction RMSE: < 0.06Generation Novelty: > 99.3%Property Prediction: Matches/surpasses

HuggingFace Causal LM

Not available in Spectra

Unconditioned de-novo molecular generation via a causal language model fine-tuned on chemical SMILES.

Published benchmarks: —

GuacaMol

Not available in Spectra

Goal-directed molecular generation benchmark suite.

Published benchmarks: —