MatGL is a universal GNN-based architecture designed to predict diverse materials properties across different crystal systems. Rather than training separate models for different properties or material classes, MatGL attempts to capture shared patterns in how atomic structure determines materials behavior.
When to use MatGL:
Screening large material databases for multiple properties simultaneously (band gap, formation energy, elastic constants, etc.)
Rapid downstream validation in generative workflows (e.g., checking if flow-generated structures have reasonable properties)
Projects where you need property predictions across diverse materials classes without retraining
Initial property estimates when you need speed over perfect accuracy
When to consider alternatives:
When you need very high accuracy for a specific property in a narrow materials class (CHGNet or task-specific models may be better)
When properties have strong long-range dependencies that GNNs struggle with
When you need physical insights into why a material has certain properties (GNNs are black boxes)
Preparation:
Structures should be reasonably relaxed (ground state or local minimum)
No special preprocessing typically needed—the model handles most variations
Works with supercells, but computation scales with system size
Expected structure sizes: Atoms per unit cell typically 1-100 (works well up to ~200 atoms)
MatGL returns property predictions in JSON format. Depending on configuration, typical outputs include:
Formation energy (eV/atom): Energy relative to elemental reference states
Band gap (eV): Electronic band gap (if trained on this property)
Elastic constants (GPa): Full elastic tensor components (if available in training)
Density of states (optional): If model was trained on DOS data
Important: MatGL outputs are point estimates. No confidence intervals or uncertainty quantification built in—you're working with single predictions.
MatGL uses a universal Materials Graph Network (M3GN) architecture:
Input: Atomic coordinates, species, and lattice vectors
Processing: Iterative message passing over nearest-neighbor graphs
Output: Per-atom or per-structure features aggregated to material properties
Training: Trained on DFT data (typically PBE functional from Materials Project or ICSD)
Compute: MatGL inference is relatively lightweight. A typical structure (~50 atoms) should complete in <1 second on CPU, milliseconds on GPU.
Batch processing: Designed for single-structure inference, but can be wrapped for batch operations on multiple materials.
Customization: The universal model is fixed, but Ouro could offer fine-tuning workflows for specific material classes or properties (Phase 2).
Validation: Ground-truth comparison against DFT for accuracy assessment is straightforward—VASP, QUANTUM ESPRESSO outputs are directly comparable.
CHGNet is a graph neural network trained to predict interatomic potentials and forces in materials. Unlike MatGL (which predicts aggregate properties), CHGNet models atomic-level forces and energies. This makes it useful for molecular dynamics simulations, structure relaxation, and understanding how atoms respond to each other.
When to use CHGNet:
Running molecular dynamics simulations without explicit DFT (much faster than DFT-based MD)
Structure relaxation and geometry optimization
Exploring configurational space via MD sampling
Studying atomic-level dynamics and ion migration
Training on structures where you want atomic-resolution physics
When to consider alternatives:
For quick property screening across many materials (use MatGL instead)
For properties that don't have strong force-field dependencies (e.g., electronic properties)
When you need sub-ångström accuracy for interatomic distances (classical MD limitations apply)
For systems where charge transfer physics is critical and not captured by the training set
Input: Atomic coordinates, species, lattice vectors (same as MatGL)
Preparation:
Structures for MD typically start from relaxed configurations
CHGNet can handle various supercell sizes (it's trained to generalize)
Works well from 10-500+ atoms per supercell
CHGNet returns physics-relevant quantities for each structure:
Total energy (eV): Energy of the structure
Atomic forces (eV/Å): Force vector for each atom (3 components)
Stress tensor (GPa): 6 independent components for lattice strain calculations
Magnetic moments (optional): Per-atom magnetization if trained on magnetic systems
Important: These are machine-learned predictions, not ground truth. Accuracy varies by material class and chemical environment. CHGNet was trained primarily on oxide systems—extrapolation to very different chemistries (e.g., intermetallics) has larger uncertainty.
CHGNet uses charge-informed graph neural networks:
Input: Atomic coordinates, species, partial charges (optional but improves accuracy)
Processing: Multi-layer message passing with explicit charge state information
Output: Per-atom energies and forces that satisfy Newton's third law and rotational invariance
Training: Trained on AIMD (DFT-based molecular dynamics) trajectories, primarily for oxides
The key innovation is incorporating charge information explicitly, which helps the model capture electronegativity effects and ionic interactions.
Compute: CHGNet is heavier than MatGL. Single-point force/energy evaluation on a 100-atom structure: ~10-50 ms on CPU, 1-5 ms on GPU. MD trajectories (100s of steps) are practical on GPU clusters.
Workflow integration: Natural fit for structure relaxation → property prediction pipelines. Relax with CHGNet, then analyze with MatGL.
Limitations to document:
Trained primarily on oxides; extrapolation to other chemistries introduces uncertainty
No explicit temperature effects (it's 0K physics)
Accuracy degrades for very high-energy configurations or unusual coordination
Validation: Compare CHGNet-relaxed structures to DFT-relaxed structures. Compare MD trajectories from CHGNet to AIMD for ion migration studies.
Aspect | MatGL | CHGNet |
|---|---|---|
Predicts | Aggregate properties | Atomic forces & energy |
Use for | Rapid screening | MD, relaxation, dynamics |
Training data | Property datasets | Trajectory data |
Output scale | Per-structure | Per-atom + global |
Speed | Very fast | Fast-to-moderate |
Accuracy focus | Diverse properties | Force/stress accuracy |
Best for oxides? | General | Excellent |
Generate candidate structures (e.g., via flow matching or genetic algorithms)
Quick MatGL screening: Predict properties (band gap, formation energy) to filter candidates
CHGNet relaxation: Relax survivors to ground state using ML-based MD
MatGL re-evaluation: Re-predict properties on relaxed structures
DFT validation: Validate most promising candidates with explicit DFT
This pipeline combines the speed and breadth of MatGL with the physical realism of CHGNet dynamics.
On this page