ORB v3

What is ORB v3?

ORB v3 is a family of universal machine-learned interatomic potentials from Orbital Materials. It maps atomic species and 3D coordinates to potential energy and force predictions, replacing repeated electronic-structure calculations with neural-network inference. Compatible checkpoints can also return stress and per-atom confidence distributions.

The model family was designed around a practical tradeoff between predictive performance, speed, memory use, and smoothness of the potential-energy surface. ORB v3 therefore includes conservative and direct force models, plus dense and sparse atomic neighborhoods. ProteinIQ runs the pinned orb-models==0.5.5 release and includes six checkpoints from that version:

• Four OMAT checkpoints for broad atomistic materials calculations
• Two OrbMol checkpoints trained on OMol25 for charge-conditioned and spin-conditioned molecular systems

ORB v3 is not a protein structure predictor or a conventional biomolecular force field. It evaluates an atomic geometry that already exists. A PDB structure can be submitted, but ORB does not infer missing atoms, protonation states, solvent, ions, restraints, or biological validity.

Three calculation modes

Mode	What it does	Best suited to	Main outputs
Single-point properties	Evaluates the submitted coordinates without moving atoms	Comparing properties for prepared geometries	Energy, forces, optional stress, confidence, evaluated structure
Geometry optimization	Relaxes coordinates with ASE BFGS until the force threshold or step limit is reached	Removing local strain and finding a nearby local minimum	Initial and final energy, convergence status, optimized structure, trajectory, step metrics
Molecular dynamics	Runs constant-volume ASE Langevin dynamics in a user-defined periodic cubic cell	Short ORB trajectory experiments with the published protocol	Final structure, thermodynamic history, extended-XYZ trajectory

Geometry optimization is the default. These modes answer different questions. Optimization finds a nearby minimum, not the global minimum. Molecular dynamics samples motion under the selected protocol, but does not establish experimental stability.

How to use ORB v3 online

Run ORB v3 online by uploading one coordinate-bearing atomic structure or fetching a PDB entry from RCSB, then selecting a single-point, geometry-optimization, or Langevin molecular-dynamics calculation. ProteinIQ returns a populated results summary, per-atom forces, downloadable structures, provenance, and mode-specific trajectories or logs.

Input

Input	Requirements
Atomic structure	One PDB, ENT, CIF, mmCIF, XYZ, or extended-XYZ file, up to 50 MB. An RCSB PDB ID can be fetched directly.

Coordinates and element identities must already be correct. For biomolecular inputs, PDB Fixer can repair missing atoms and residues, while MolProbity can identify clashes and geometry outliers before an ORB calculation.

The PDB formatting review shown during upload is a ProteinIQ file-quality check, not an ORB failure. Warning-only records remain unchanged unless an explicit formatting action is selected.

Choose the calculation mode first

• Single point: Preserves the input geometry. Use it when coordinates have already been prepared and only the predicted properties are needed.
• Geometry optimization: Changes the coordinates. Use it to relax local strain or prepare a structure for a same-model energy comparison.
• Molecular dynamics: Requires an explicit cubic cell size and makes the system periodic in all three dimensions. Use it only when those boundary conditions are scientifically appropriate.

Choose an ORB model

Model choice	Scientific behavior	When it fits
Conservative OMAT, 120 neighbors	Forces are energy gradients and the dense neighborhood gives the smoothest OMAT option in this tool	Recommended starting point for single-point properties and optimization
Conservative OMAT, 20 neighbors	Energy-derived forces with a sparse neighborhood	Lower memory use when the 20-neighbor discontinuities are acceptable
Direct OMAT, 120 neighbors	Predicts forces directly with a dense neighborhood	Faster large-system work where strict energy conservation is not required
Direct OMAT, 20 neighbors	Fastest, sparsest OMAT checkpoint; also used by the published ORB MD example	Short source-example MD runs and throughput-sensitive calculations
Conservative OrbMol	Molecular checkpoint with energy-derived forces and explicit charge and spin conditioning	Molecular single-point or optimization calculations where smoothness matters
Direct OrbMol	Faster molecular checkpoint with directly predicted forces and explicit charge and spin conditioning	Throughput-sensitive molecular single-point or optimization calculations

The 120 neighbors labels correspond to the effectively unlimited inf checkpoints in the ORB model names. The 20-neighbor models reduce graph processing but introduce neighborhood discontinuities. The ORB v3 paper recommends conservative inf models for calculations that require rigorous energy conservation, especially constant-energy dynamics.

OMAT models were trained on the AIMD subset of OMat24 and can return stress. OrbMol models were trained on OMol25, require explicit Total charge and Total spin, and do not return stress. Molecular dynamics uses OMAT only.

Shared settings

Setting	Description
Calculation mode	Selects single-point properties, BFGS geometry optimization, or Langevin NVT molecular dynamics. Default: Geometry optimization.
Model variant	Selects an OMAT or OrbMol checkpoint for single-point and optimization runs. Default: Conservative OMAT, 120 neighbors.
MD model variant	Selects one of the four OMAT checkpoints for dynamics. Default: Direct OMAT, 20 neighbors, matching the published example.
Precision	ORB loader precision: Float32 high (default), Float32 highest, or Float64. Higher precision increases memory and compute requirements.
Total charge	Required explicit conditioning value for OrbMol. No neutral default is assumed.
Total spin	Required explicit conditioning value for OrbMol. No zero-spin default is assumed.

Geometry-optimization settings

Setting	Description
Convergence threshold (fmax)	ASE BFGS stops when the largest per-atom force norm is at or below this value. Range: 0.001 to 0.1 eV/Å. Default: 0.01 eV/Å.
Maximum optimization steps	Stop limit from 1 to 5,000 steps. Default: 500. The optimizer can finish earlier after reaching fmax.

Molecular-dynamics settings

Setting	Description
Cubic cell size	Required periodic cell length in Å. The ORB NaCl and water example uses 25.25, but arbitrary inputs need a cell selected for their own dimensions.
Temperature	Langevin thermostat target. Default: 300 K.
Timestep	Integration timestep. Default: 0.5 fs.
Friction	Langevin friction coefficient. Default: 0.01/fs.
MD steps	Number of integration steps. Default: 100. Maximum: 5,000.
Trajectory interval	Steps between stored extended-XYZ coordinate frames. Default: 20.

The ASE thermodynamic log records every step, as in the ORB example.

Initial velocities are sampled with ASE MaxwellBoltzmannDistribution. No random seed control is added, so repeated MD jobs are expected to produce different trajectories.

Understanding ORB v3 results

For a successful job, the Results tab opens first and contains the final summary row. Run history appears when optimization or molecular dynamics returns stepwise data. Structure previews the mode-specific PDB derivative, and Files contains every stored artifact.

Results table

Result	Interpretation
Energy (eV)	Potential energy for a single-point geometry.
Initial energy (eV)	Energy before optimization or dynamics.
Final energy (eV)	Energy of the final coordinates.
Energy change (eV)	Final minus initial energy. A negative value means the final geometry is lower under the same model.
Maximum force (eV/Å)	Largest norm among the final per-atom force vectors. Optimization compares this value with fmax.
Stress returned	Whether ORB supplied a finite six-component stress result. Absence is reported rather than replaced with a fabricated value.
Confidence returned	Whether the checkpoint supplied per-atom force-error confidence bins.
Optimization steps	Number of BFGS steps completed.
Converged	true only when BFGS stopped successfully and the final maximum force is at or below fmax.
Final temperature (K)	ASE temperature calculated from the final MD state. It will fluctuate around the thermostat target.

Energy comparisons are valid only when atom composition, checkpoint, charge and spin conditioning, numerical precision, and boundary conditions are compatible. Absolute energies from different model families or differently composed systems are not directly rankable.

Downloadable files

File	Description
evaluated_structure.extxyz, optimized_structure.extxyz, or final_structure.extxyz	Authoritative full-precision ASE coordinates for the selected mode.
Matching .pdb structure	Rounded derivative used for 3D preview and structure workflows. It is not the authoritative coordinate file.
per_atom_forces.csv	Atom index, element, Cartesian force components, and force-vector norm in eV/Å.
stress_tensor.csv	Finite stress in ASE Voigt order: xx, yy, zz, yz, xz, xy, in eV/ų. Returned only when available.
per_atom_confidence.csv	Predicted force-error bin, bin boundaries, and all confidence-bin probabilities for each atom. Returned only when available.
orb_summary.json	Machine-readable settings, scalar results, warnings, package version, source commit, and checkpoint provenance.

Optimization also returns:

File	Description
optimization_trajectory.extxyz	Full-precision coordinate frames recorded during ASE BFGS.
optimization_metrics.csv	Step, energy, and maximum force for the optimization history.
optimization.log	Native ASE BFGS log.

Molecular dynamics also returns:

File	Description
md_trajectory.extxyz	Periodic Langevin trajectory written at the selected interval.
md_nvt.log	Native ASE MDLogger table with time, total energy, potential energy, kinetic energy, and temperature.

Reading confidence output

ORB confidence is a per-atom distribution over predicted force-error intervals. The predicted_bin is the interval with the highest probability, while the probability columns retain the complete distribution. Narrow, lower-error bins with concentrated probability indicate greater model confidence than broad or high-error bins.

Confidence is not experimental validation and is not a calibrated probability that an entire structure is correct. Inspect high-error or diffuse atoms in structural context, especially around unusual chemistry, strained contacts, metals, defects, surfaces, or charged groups.

Reading optimization and MD histories

For optimization, a decreasing energy and maximum force usually indicate progress toward a local minimum. A job that reaches the step limit is still returned with its final properties and complete history, but Converged remains false.

For MD, inspect temperature and total-energy behavior together with the trajectory. Sudden energy spikes, unrealistic atomic separations, or rapid structural disruption can indicate unsuitable starting coordinates, a timestep that is too large, poor boundary conditions, or chemistry outside the checkpoint's reliable domain.

How ORB v3 works

• Atomic graph construction: ORB converts elements, coordinates, periodic boundaries, and model conditioning into an atomic graph.
• Neural potential inference: The selected checkpoint predicts total energy, per-atom forces, and optional stress and confidence. Conservative checkpoints obtain forces from the energy gradient; direct checkpoints predict forces separately.
• Coordinate propagation: Single-point mode stops after property evaluation. Optimization passes ORB forces to ASE BFGS. Molecular dynamics passes them to ASE Langevin dynamics with Maxwell-Boltzmann initial velocities.
• Final evaluation and export: ORB recalculates properties for the final coordinates. ProteinIQ stores full-precision extended XYZ, an additive PDB preview, tabular results, native logs, and exact source provenance.

When to use ORB v3 vs alternatives

ORB v3 is a strong fit for materials-oriented energy and force calculations, rapid local relaxation, and short atomistic simulations where a universal learned potential is appropriate. Its direct models are especially useful when system size or throughput would make a conservative potential too expensive.

For isolated organic molecules where atomic charges, dipoles, Hessians, ensemble uncertainty, or specialized elemental coverage matter, compare the same prepared geometry with AIMNet2. For protein or protein-ligand dynamics that require protonation, solvent construction, established biomolecular force fields, constraints, pressure control, and longer production protocols, use OpenMM.

Limitations

• Predictions depend on the chemical and structural coverage learned by the selected checkpoint. Unusual bonding, metal coordination, highly charged systems, defects, and strongly out-of-distribution structures need independent validation.
• OMAT models were trained on materials data. The ORB authors report transfer to non-periodic molecular and biomolecular systems, but that does not make them a drop-in replacement for a validated domain-specific method.
• OrbMol requires scientifically correct total charge and total spin. ProteinIQ cannot infer them safely from coordinates alone.
• Stress may be undefined or non-finite for non-periodic structures. ProteinIQ omits the tensor file and reports stress as unavailable in that case.
• A 20-neighbor cutoff introduces discontinuities in the potential-energy surface. Direct models are also non-conservative because forces are not constrained to equal the energy gradient.
• PDB files often omit hydrogens, solvent, ions, cofactors, alternate conformations, and protonation details. ORB does not repair or parameterize these features.
• The MD mode is the published constant-volume Langevin example with a hosted 5,000-step limit. It does not add solvent, neutralize the system, apply bond constraints, run equilibration stages, control pressure, or provide a turnkey biomolecular simulation protocol.