Dispersion correction cutoff parsing

[yjw0510]

In foundation_models.py, the default value for the argument dispersion_cutoff is set to 40.0 * units.Bohr
mace/mace/calculators/foundations_models.py

def mace_mp(
    model: Union[str, Path] = None,
    device: str = "",
    default_dtype: str = "float32",
    dispersion: bool = False,
    damping: Literal["zero", "bj", "zerom", "bjm"] = "bj",
    dispersion_xc: str = "pbe",
    dispersion_cutoff: float = 40.0 * units.Bohr,
    **kwargs,
) -> MACECalculator:

Later in the code, we can see that

        d3_calc = TorchDFTD3Calculator(
            device=device,
            damping=damping,
            dtype=dtype,
            xc=dispersion_xc,
            cutoff=dispersion_cutoff,
            **kwargs,
        )

However, when I tracked the passed argument in torch-dftd, I found that the default value for the cutoff argument is set to 95.0 * Bohr, while the cnthr argument is set to 40.0 * Bohr.

torch-dftd/torch_dftd/torch_dftd3_calculator.py

class TorchDFTD3Calculator(Calculator):
...
    def __init__(
        self,
        dft: Optional[Calculator] = None,
        atoms: Atoms = None,
        damping: str = "zero",
        xc: str = "pbe",
        old: bool = False,
        device: str = "cpu",
        cutoff: float = 95.0 * Bohr,
        cnthr: float = 40.0 * Bohr,
        abc: bool = False,
        # --- torch dftd3 specific params ---
        dtype: torch.dtype = torch.float32,
        bidirectional: bool = True,
        cutoff_smoothing: str = "none",
        **kwargs,
    ):

Is this simply a mistake confusing cutoff for cnthr, or is it an intended implementation to decrease the computational load coming from the large cutoff value in the dispersion correction?

[ilyes319]

Hey @yjw0510,

Indeed we reduced the cutoff to speed up the computation. We made some limited experiment to select the new default in order to compromise the least the accuracy. More clever approaches could be taken, like adding on top a longer 2 body dispersion, but we have not yet added that.