Choosing a calculator

Atomli calculators share ASE’s attachment model, but the native backends do more than provide a drop-in method name. They retain expensive state between requests, use the same compiled engines in optimization and dynamics, and enforce each method’s scientific boundary before a calculation starts.

Choose by workload

Calculator	Best fit	Python properties	Periodic systems	External data
QC	Molecular and periodic PBE or r2SCAN calculations	energy, forces, periodic stress	Three-dimensional PBC	Packaged basis and pseudopotential data
XTB	Fast electronic-structure energies and relaxation	energy, forces, stress	Yes	None
MLIP	Large structures, long trajectories, materials screening	energy, forces, stress	Yes, when the model supports the species	Model weights downloaded once or supplied locally
DFTD3	Standalone dispersion correction	energy, forces	Yes	None
LennardJones	Tests and simple classical pair models	energy, forces	No cell-specific stress in Python	None

Choose by method

• PBE: default GGA DFT for molecular and periodic systems when SCF cost is acceptable.
• r2SCAN: meta-GGA DFT when energetics justify the extra cost over PBE.
• SKALA 1.1: neural exchange-correlation functional for molecular calculations, with automatic checkpoint delivery.
• GFN2-xTB: fast tight binding for relaxation, screening, and charged or periodic systems.
• g-xTB: tight binding with higher target accuracy than GFN2-xTB at comparable cost.
• MLIP models: Nequix, NequIP, and Equiformer potentials for large systems and long trajectories on CPU or WGPU.
• DFT-D3: standalone dispersion correction to combine with PBE or r2SCAN.

PBE, r2SCAN, and SKALA 1.1 can evaluate the SCF on a GPU for closed-shell molecules.

Energy and forces

from atomli import Atoms
from atomli.calculators.xtb import XTB

atoms = Atoms("H2", positions=[[0, 0, 0], [0, 0, 0.74]])
atoms.calc = XTB(method="gfn2")

energy = atoms.get_potential_energy()
forces = atoms.get_forces()

The calculator is retained on the Atoms object. A force request immediately after an energy request can reuse backend state instead of reconstructing the native engine.

Python API coverage

The calculator guides separate public Python features from native features without Python bindings:

• Python features means the current public Atomli wheel exposes the constructor, property, or behavior.
• Native-only features means the implementation exists in qc.rs, tb.rs, mlip.rs, or their ASE adapter, but its higher-level Python binding is still future work.

ASE interoperability

Atomli calculators can be attached directly to ase.Atoms:

from ase import Atoms
from atomli.calculators.xtb import XTB

atoms = Atoms("H2", positions=[[0, 0, 0], [0, 0, 0.74]])
atoms.calc = XTB(method="gfn2")

ASE calculators assigned to Atomli Atoms are adapted on assignment. No user-facing bridge or private-module import is required.

Unsupported properties

Each calculator publishes implemented_properties. Requesting an unsupported property raises NotImplementedError or a calculation error. Atomli does not silently replace a requested method, device, or property with a different one.