One of the most common bottlenecks in computational materials science is getting a quick property estimate for a crystal structure without running a full DFT calculation. ALIGNN (Atomistic Line Graph Neural Network) solves this by providing pretrained models that predict dozens of materials properties directly from structure — and now all of them are available as individual API endpoints on Ouro.
Upload a CIF file, pick the property you care about, get a prediction back in seconds.
Access all ALIGNN pretrained models through individual endpoints. Each endpoint accepts a CIF file and returns a JSON prediction. Models span energetics, electronic structure, mechanical properties, thermoelectrics, superconductivity, magnetism, dielectrics, catalysis, MOFs, and molecular properties.
The API exposes 50+ pretrained ALIGNN models spanning nearly every major property domain in materials science:
Energetics & stability: formation energy, total energy, energy above the convex hull (JARVIS-DFT and Materials Project)
Electronic structure: band gaps (optB88vdW, TBmBJ, MP-PBE), band edge positions (CBM, VBM), work function, effective masses
Mechanical properties: bulk modulus, shear modulus, exfoliation energy
Dielectrics & piezoelectrics: static and electronic dielectric functions, maximum piezoelectric strain coefficient
Thermoelectrics: Seebeck coefficients (n-type and p-type), thermoelectric power factor
Superconductivity: critical temperature, electronic DOS at Fermi level, Debye temperature, Eliashberg spectral function
Magnetism: total magnetic moment, electric field gradient
Catalysis & surfaces: adsorption energies from TinNet, AGRA, and Open Catalyst Project datasets
MOF properties: CO adsorption, surface area, pore/cavity diameters, void fraction
Molecular properties: QM9 targets including internal energy, HOMO-LUMO gap, polarizability, ZPVE
Solar cells: spectroscopic limited maximum efficiency (SLME)
Topological materials: spin-orbit spillage
Spectral predictions: phonon density of states
Every model is its own route. You don't need to configure anything — just send a CIF to the right endpoint.
Speed: get property predictions in seconds instead of hours or days of DFT compute.
Breadth: screen the same structure across many properties without setting up separate workflows for each one.
Consistency: all models use the same ALIGNN architecture trained on curated datasets (primarily JARVIS-DFT), so predictions are methodologically comparable.
Simplicity: every endpoint has the same interface — POST a CIF file, get back JSON.
This is useful whether you're triaging candidates from a generative model, building a screening pipeline, or just want a quick sanity check on a structure before committing to more expensive calculations.
Predict formation energy per atom:
Predicts the DFT formation energy per atom from the JARVIS-DFT optB88vdW dataset.
Predict band gap (TBmBJ functional):
Predicts the electronic band gap from the JARVIS-DFT TBmBJ functional. Generally more accurate than PBE/optB88vdW for band gaps.
Predict superconducting critical temperature:
Predicts the superconducting critical temperature Tc.
Or list every available model with a single GET request:
ALIGNN represents crystal structures as line graphs — a graph where both atoms and bonds are treated as nodes, capturing angular and many-body interactions that simpler GNN architectures miss. Each pretrained model was trained on a specific property target from a curated dataset (JARVIS-DFT, Materials Project, QM9, Open Catalyst, etc.).
When you call an endpoint:
Your CIF file is parsed into an atomic structure.
The ALIGNN line graph is constructed from the geometry.
The appropriate pretrained model runs inference.
The predicted property value is returned as JSON.
The underlying models and training data come from the JARVIS-Tools / ALIGNN project at NIST.
ALIGNN predictions are fast enough to use as filters in large-scale workflows. Some practical patterns:
Generative model post-processing: run stability and property checks on every structure a generative model outputs, before deciding which ones to relax or synthesize.
Multi-property screening: for a given candidate, quickly check formation energy, band gap, mechanical stability, and functional properties in parallel.
Dataset enrichment: augment an existing materials dataset with predicted properties that weren't originally computed.
Research exploration: when reading a paper about a new compound, quickly check what ALIGNN predicts for it across multiple property axes.
These are ML predictions, not DFT calculations.
Accuracy varies by model and by how similar your structure is to the training data. Models trained on JARVIS-DFT are most reliable for bulk inorganic crystals.
Band gaps from optB88vdW are known to underestimate; TBmBJ predictions are generally more accurate.
MOF and molecular (QM9) models are trained on specific structure types and may not transfer well outside those domains.
For any result that matters, treat the prediction as a screening signal and validate with higher-fidelity methods.
If you have questions or feedback on which models or properties to add next, let us know.
On this page
50+ pretrained graph neural network models for predicting materials properties from a CIF file. Covers energetics, band gaps, mechanical properties, thermoelectrics, superconductivity, catalysis, MOFs, and more.