pycontrails 0.58.0__cp314-cp314-win_amd64.whl

pycontrails/core/interpolation.py

@@ -0,0 +1,702 @@
+"""Interpolation utilities."""
+
+from __future__ import annotations
+
+import dataclasses
+import logging
+from typing import Any, Literal, overload
+
+import numpy as np
+import numpy.typing as npt
+import scipy.interpolate
+import xarray as xr
+
+from pycontrails.core import rgi_cython  # type: ignore[attr-defined]
+
+# ------------------------------------------------------------------------------
+# Multidimensional interpolation
+# ------------------------------------------------------------------------------
+
+
+logger = logging.getLogger(__name__)
+
+
+class PycontrailsRegularGridInterpolator(scipy.interpolate.RegularGridInterpolator):
+    """Support for performant interpolation over a regular grid.
+
+    This class is a thin wrapper around the
+    :class:`scipy.interpolate.RegularGridInterpolator` in order to make typical
+    ``pycontrails`` linear interpolation use-cases more performant:
+
+    #. Avoid ``RegularGridInterpolator`` constructor validation when `method="linear"`.
+       In :func:`interp`, parameters are carefully crafted to fit into the intended form,
+       thereby making validation unnecessary.
+    #. Override the :meth:`_evaluate_linear` method with a faster implementation. See
+       the :meth:`_evaluate_linear` docstring for more information.
+
+    **This class should not be used directly. Instead, use the ``interp`` function.**
+
+    .. versionchanged:: 0.40.0
+
+        The :meth:`_evaluate_linear` method now uses a Cython implementation. The dtype
+        of the output is now consistent with the dtype of the underlying :attr:`values`
+
+    .. versionchanged:: 0.58.0
+
+        Any ``method`` other than ``"linear"`` now uses the
+        :class:`scipy.interpolate.RegularGridInterpolator` implementation. This
+        allows for greater flexibility in the ``method`` parameter.
+
+    Parameters
+    ----------
+    points : tuple[npt.NDArray[np.floating], ...]
+        Coordinates of the grid points.
+    values : npt.NDArray[np.floating]
+        Grid values. The shape of this array must be compatible with the
+        coordinates.
+    method : str
+        Passed into :class:`scipy.interpolate.RegularGridInterpolator`
+    bounds_error : bool
+        Passed into :class:`scipy.interpolate.RegularGridInterpolator`
+    fill_value : float | np.float64 | None
+        Passed into :class:`scipy.interpolate.RegularGridInterpolator`
+
+    See Also
+    --------
+    scipy.interpolate.RegularGridInterpolator
+    interp
+    """
+
+    def __init__(
+        self,
+        points: tuple[npt.NDArray[np.floating], ...],
+        values: npt.NDArray[np.floating],
+        *,
+        method: str,
+        bounds_error: bool,
+        fill_value: float | np.float64 | None,
+    ) -> None:
+        if method != "linear" or values.dtype not in (np.float32, np.float64):
+            # Slow path: use parent class
+            super().__init__(
+                points,
+                values,
+                method=method,
+                bounds_error=bounds_error,
+                fill_value=fill_value,
+            )
+            return
+
+        # Fast path: no validation
+        self.grid = points
+        self.values = values
+        self.method = method
+        self.bounds_error = bounds_error
+        self.fill_value = fill_value
+        self._spline = None  # XXX: setting private attribute on RGI
+
+    def _prepare_xi_simple(self, xi: npt.NDArray[np.floating]) -> npt.NDArray[np.bool_]:
+        """Run looser version of :meth:`_prepare_xi`.
+
+        Parameters
+        ----------
+        xi : npt.NDArray[np.floating]
+            Points at which to interpolate.
+
+        Returns
+        -------
+        npt.NDArray[np.bool_]
+            A 1-dimensional Boolean array indicating which points are out of bounds.
+            If ``bounds_error`` is ``True``, this will be all ``False``.
+        """
+
+        if self.bounds_error:
+            for i, p in enumerate(xi.T):
+                g0 = self.grid[i][0]
+                g1 = self.grid[i][-1]
+                if not (np.all(p >= g0) and np.all(p <= g1)):
+                    msg = f"One of the requested xi is out of bounds in dimension {i}"
+                    raise ValueError(msg)
+
+            return np.zeros(xi.shape[0], dtype=bool)
+
+        return self._find_out_of_bounds(xi.T)  # XXX: calling private method on RGI
+
+    def __call__(
+        self, xi: npt.NDArray[np.floating], method: str | None = None
+    ) -> npt.NDArray[np.floating]:
+        """Evaluate the interpolator at the given points.
+
+        Parameters
+        ----------
+        xi : npt.NDArray[np.floating]
+            Points at which to interpolate. Must have shape ``(n, ndim)``, where
+            ``ndim`` is the number of dimensions of the interpolator.
+        method : str | None
+            Override the :attr:`method` to keep parity with the base class.
+
+        Returns
+        -------
+        npt.NDArray[np.floating]
+            Interpolated values. Has shape ``(n,)``. When computing linear interpolation,
+            the dtype is the same as the :attr:`values` array.
+        """
+
+        method = method or self.method
+        if method != "linear":
+            return super().__call__(xi, method)
+
+        out_of_bounds = self._prepare_xi_simple(xi)
+        xi_indices, norm_distances = self._find_indices(xi.T)  # XXX: calling private method on RGI
+
+        out = self._evaluate_linear(xi_indices, norm_distances)
+        return self._set_out_of_bounds(out, out_of_bounds)
+
+    def _set_out_of_bounds(
+        self,
+        out: npt.NDArray[np.floating],
+        out_of_bounds: npt.NDArray[np.bool_],
+    ) -> npt.NDArray[np.floating]:
+        """Set out-of-bounds values to the fill value.
+
+        Parameters
+        ----------
+        out : npt.NDArray[np.floating]
+            Values from interpolation. This is modified in-place.
+        out_of_bounds : npt.NDArray[np.bool_]
+            A 1-dimensional Boolean array indicating which points are out of bounds.
+
+        Returns
+        -------
+        out : npt.NDArray[np.floating]
+            A reference to the ``out`` array.
+        """
+        if self.fill_value is not None and np.any(out_of_bounds):
+            out[out_of_bounds] = self.fill_value
+
+        return out
+
+    def _evaluate_linear(
+        self,
+        indices: npt.NDArray[np.int64],
+        norm_distances: npt.NDArray[np.floating],
+    ) -> npt.NDArray[np.floating]:
+        """Evaluate the interpolator using linear interpolation.
+
+        This is a faster alternative to
+        :meth:`scipy.interpolate.RegularGridInterpolator._evaluate_linear`.
+
+        .. versionadded:: 0.24
+
+        .. versionchanged:: 0.40.0
+
+            Use Cython routines for evaluating the interpolation when the
+            dimension is 1, 2, 3, or 4.
+
+        Parameters
+        ----------
+        indices : npt.NDArray[np.int64]
+            Indices of the grid points to the left of the interpolation points.
+            Has shape ``(ndim, n_points)``.
+        norm_distances : npt.NDArray[np.floating]
+            Normalized distances between the interpolation points and the grid
+            points to the left. Has shape ``(ndim, n_points)``.
+
+        Returns
+        -------
+        npt.NDArray[np.floating]
+            Interpolated values with shape ``(n_points,)`` and the same dtype as
+            the :attr:`values` array.
+        """
+        # Let scipy "slow" implementation deal with high-dimensional grids
+        if indices.shape[0] > 4:
+            return super()._evaluate_linear(indices, norm_distances)
+
+        # Squeeze as much as possible
+        # Our cython implementation requires non-degenerate arrays
+        non_degen = tuple(s > 1 for s in self.values.shape)
+        values = self.values.squeeze()
+        indices = indices[non_degen, :]
+        norm_distances = norm_distances[non_degen, :]
+
+        ndim, n_points = indices.shape
+        out = np.empty(n_points, dtype=self.values.dtype)
+
+        if ndim == 4:
+            return rgi_cython.evaluate_linear_4d(values, indices, norm_distances, out)
+
+        if ndim == 3:
+            return rgi_cython.evaluate_linear_3d(values, indices, norm_distances, out)
+
+        if ndim == 2:
+            return rgi_cython.evaluate_linear_2d(values, indices, norm_distances, out)
+
+        if ndim == 1:
+            # np.interp could be better ... although that may also promote the dtype
+            # 1-d view is required for evaluate_linear_1d
+            return rgi_cython.evaluate_linear_1d(values, indices[0, :], norm_distances[0, :], out)
+
+        msg = f"Invalid number of dimensions: {ndim}"
+        raise ValueError(msg)
+
+
+def _floatize_time(
+    time: npt.NDArray[np.datetime64], offset: np.datetime64
+) -> npt.NDArray[np.floating]:
+    """Convert an array of ``np.datetime64`` to an array of ``np.float64``.
+
+    In calls to :class:`scipy.interpolate.RegularGridInterpolator`, it's critical
+    that every coordinate be of same type. This creates complications: spatial
+    coordinates are float-like, whereas time coordinates are datetime-like. In
+    particular, it is not possible to cast an ``np.datetime64`` to a float
+    without losing information. In practice, this is not problematic because
+    ``np.float64`` has plenty of precision. Previously, this was more of an issue
+    because we used ``np.float32``.
+
+    This function uses a fixed time resolution (1 millisecond) to convert the time
+    coordinate to a float-like coordinate. The time resolution is taken to avoid
+    losing too much information for the time scales we encounter.
+
+    Care is taken to ensure "nat" values are converted to "nan".
+
+    Note that ``xarray`` also must confront this issue. They take a similar approach
+    in :func:`xarray.core.missing._floatize_x`. See
+    https://github.com/pydata/xarray/blob/d4db16699f30ad1dc3e6861601247abf4ac96567/xarray/core/missing.py#L572
+
+    .. versionchanged:: 0.40.0
+
+        No longer allow the option of converting to ``np.float32``. No longer
+        floor the time values to the preceding millisecond.
+
+    Parameters
+    ----------
+    time : npt.NDArray[np.datetime64]
+        Array of ``np.datetime64`` values.
+    offset : np.datetime64
+        The offset to subtract from ``time``.
+
+    Returns
+    -------
+    npt.NDArray[np.floating]
+        The number of milliseconds since ``offset``.
+    """
+    delta = time - offset
+    resolution = np.timedelta64(1, "ms")
+    return delta / resolution
+
+
+def _localize(da: xr.DataArray, coords: dict[str, np.ndarray]) -> xr.DataArray:
+    """Clip ``da`` to the smallest bounding box that contains all of ``coords``.
+
+    Roughly follows approach taken by :func:`xarray.core.missing._localize`. See
+    https://github.com/pydata/xarray/blob/56f05c37924071eb4712479d47432aafd4dce38b/xarray/core/missing.py#L557
+
+    Parameters
+    ----------
+    da : xr.DataArray
+        DataArray to clip.
+    coords : dict[str, np.ndarray]
+        Coordinates to clip to.
+
+    Returns
+    -------
+    xr.DataArray
+        Clipped :class:`xarray.DataArray`. Has the same dimensions as the input ``da``.
+        In particular, each dimension of the returned DataArray is a slice of the
+        corresponding dimension of the input ``da``.
+    """
+    indexes: dict[str, Any] = {}
+    for dim, arr in coords.items():
+        dim_vals = da[dim].values
+
+        # Skip single level
+        if dim == "level" and dim_vals.size == 1 and dim_vals.item() == -1:
+            continue
+
+        # Create slice
+        minval = np.nanmin(arr)
+        maxval = np.nanmax(arr)
+        imin = np.searchsorted(dim_vals, minval, side="right") - 1
+        imin = max(0, imin)
+        imax = np.searchsorted(dim_vals, maxval, side="left") + 1
+        indexes[dim] = slice(imin, imax)
+
+        # Logging
+        n_in_bounds = np.sum((arr >= minval) & (arr <= maxval))
+        logger.debug(
+            "Interpolation in bounds along dimension %s: %d/%d",
+            dim,
+            n_in_bounds,
+            arr.size,
+        )
+
+    return da.isel(**indexes)
+
+
+@overload
+def interp(
+    longitude: npt.NDArray[np.floating],
+    latitude: npt.NDArray[np.floating],
+    level: npt.NDArray[np.floating],
+    time: npt.NDArray[np.datetime64],
+    da: xr.DataArray,
+    method: str,
+    bounds_error: bool,
+    fill_value: float | np.float64 | None,
+    localize: bool,
+    *,
+    indices: RGIArtifacts | None = ...,
+    return_indices: Literal[False] = ...,
+) -> npt.NDArray[np.floating]: ...
+
+
+@overload
+def interp(
+    longitude: npt.NDArray[np.floating],
+    latitude: npt.NDArray[np.floating],
+    level: npt.NDArray[np.floating],
+    time: npt.NDArray[np.datetime64],
+    da: xr.DataArray,
+    method: str,
+    bounds_error: bool,
+    fill_value: float | np.float64 | None,
+    localize: bool,
+    *,
+    indices: RGIArtifacts | None = ...,
+    return_indices: Literal[True],
+) -> tuple[npt.NDArray[np.floating], RGIArtifacts]: ...
+
+
+@overload
+def interp(
+    longitude: npt.NDArray[np.floating],
+    latitude: npt.NDArray[np.floating],
+    level: npt.NDArray[np.floating],
+    time: npt.NDArray[np.datetime64],
+    da: xr.DataArray,
+    method: str,
+    bounds_error: bool,
+    fill_value: float | np.float64 | None,
+    localize: bool,
+    *,
+    indices: RGIArtifacts | None = ...,
+    return_indices: bool = ...,
+) -> npt.NDArray[np.floating] | tuple[npt.NDArray[np.floating], RGIArtifacts]: ...
+
+
+def interp(
+    longitude: npt.NDArray[np.floating],
+    latitude: npt.NDArray[np.floating],
+    level: npt.NDArray[np.floating],
+    time: npt.NDArray[np.datetime64],
+    da: xr.DataArray,
+    method: str,
+    bounds_error: bool,
+    fill_value: float | np.float64 | None,
+    localize: bool,
+    *,
+    indices: RGIArtifacts | None = None,
+    return_indices: bool = False,
+) -> npt.NDArray[np.floating] | tuple[npt.NDArray[np.floating], RGIArtifacts]:
+    """Interpolate over a grid with ``localize`` option.
+
+    .. versionchanged:: 0.25.6
+
+        Utilize scipy 1.9 upgrades to remove singleton dimensions.
+
+    .. versionchanged:: 0.26.0
+
+        Include ``indices`` and ``return_indices`` experimental parameters.
+        Currently, nan values in ``longitude``, ``latitude``, ``level``, or ``time``
+        are always propagated through to the output, regardless of ``bounds_error``.
+        In other words, a ``ValueError`` for an out of bounds coordinate is only raised
+        if a non-nan value is out of bounds.
+
+    .. versionchanged:: 0.40.0
+
+        When ``return_indices`` is True, an instance of :class:`RGIArtifacts`
+        is used to store the indices artifacts.
+
+    Parameters
+    ----------
+    longitude, latitude, level, time : np.ndarray
+        Coordinates of points to be interpolated. These parameters have the same
+        meaning as ``x`` in analogy with :func:`numpy.interp`. All four of these
+        arrays must be 1 dimensional of the same size.
+    da : xr.DataArray
+        Gridded data interpolated over. Must adhere to ``MetDataArray`` conventions.
+        In particular, the dimensions of ``da`` must be ``longitude``, ``latitude``,
+        ``level``, and ``time``. The three spatial dimensions must be monotonically
+        increasing with ``float64`` dtype. The ``time`` dimension must be
+        monotonically increasing with ``datetime64`` dtype.
+        Assumed to be cheap to load into memory (:attr:`xr.DataArray.values` is
+        used without hesitation).
+    method : str
+        Passed into :class:`scipy.interpolate.RegularGridInterpolator`.
+    bounds_error : bool
+        Passed into :class:`scipy.interpolate.RegularGridInterpolator`.
+    fill_value : float | np.float64 | None
+        Passed into :class:`scipy.interpolate.RegularGridInterpolator`.
+    localize : bool
+        If True, clip ``da`` to the smallest bounding box that contains all of
+        ``coords``.
+    indices : tuple | None, optional
+        Experimental. Provide intermediate artifacts computed by
+        :meth:``scipy.interpolate.RegularGridInterpolator._find_indices`
+        to avoid redundant computation. If known and provided, this can speed
+        up interpolation by avoiding an unnecessary call to ``_find_indices``.
+        By default, None. Must be used precisely.
+    return_indices : bool, optional
+        If True, return output of :meth:`scipy.interpolate.RegularGridInterpolator._find_indices`
+        in addition to interpolated values.
+
+    Returns
+    -------
+    npt.NDArray[np.floating] | tuple[npt.NDArray[np.floating], RGIArtifacts]
+        Interpolated values with same size as ``longitude``. If ``return_indices``
+        is True, return intermediate indices artifacts as well.
+
+    See Also
+    --------
+    pycontrails.MetDataArray.interpolate
+    scipy.interpolate.interpn
+    scipy.interpolate.RegularGridInterpolator
+    """
+    if localize:
+        coords = {"longitude": longitude, "latitude": latitude, "level": level, "time": time}
+        da = _localize(da, coords)
+
+    indexes = da._indexes
+    x = indexes["longitude"].index.values  # type: ignore[attr-defined]
+    y = indexes["latitude"].index.values  # type: ignore[attr-defined]
+    z = indexes["level"].index.values  # type: ignore[attr-defined]
+    if any(v.dtype != np.float64 for v in (x, y, z)):
+        msg = "da must have float64 dtype for longitude, latitude, and level coordinates"
+        raise ValueError(msg)
+
+    # Convert t and time to float64
+    t = indexes["time"].index.values  # type: ignore[attr-defined]
+    offset = t[0]
+    t = _floatize_time(t, offset)
+
+    single_level = z.size == 1 and z.item() == -1.0
+    points: tuple[npt.NDArray[np.floating], ...]
+    if single_level:
+        values = da.values.squeeze(axis=2)
+        points = x, y, t
+    else:
+        values = da.values
+        points = x, y, z, t
+
+    interp_ = PycontrailsRegularGridInterpolator(
+        points=points,
+        values=values,
+        method=method,
+        bounds_error=bounds_error,
+        fill_value=fill_value,
+    )
+
+    if indices is None:
+        xi = _buildxi(longitude, latitude, level, time, offset, single_level)
+        if return_indices:
+            out, indices = _linear_interp_with_indices(interp_, xi, localize, None)
+            return out, indices
+        return interp_(xi)
+
+    out, indices = _linear_interp_with_indices(interp_, None, localize, indices)
+    if return_indices:
+        return out, indices
+    return out
+
+
+def _buildxi(
+    longitude: npt.NDArray[np.floating],
+    latitude: npt.NDArray[np.floating],
+    level: npt.NDArray[np.floating],
+    time: npt.NDArray[np.datetime64],
+    offset: np.datetime64,
+    single_level: bool,
+) -> npt.NDArray[np.floating]:
+    """Build the input array for interpolation.
+
+    The implementation below achieves the same result as the following::
+
+        np.stack([longitude, latitude, level, time_float], axis=1])
+
+    This implementation is slightly faster than the above.
+    """
+
+    time_float = _floatize_time(time, offset)
+
+    ndim = 3 if single_level else 4
+    shape = longitude.size, ndim
+    xi = np.empty(shape, dtype=np.float64)
+
+    xi[:, 0] = longitude
+    xi[:, 1] = latitude
+    if not single_level:
+        xi[:, 2] = level
+    xi[:, -1] = time_float
+
+    return xi
+
+
+def _linear_interp_with_indices(
+    interp: PycontrailsRegularGridInterpolator,
+    xi: npt.NDArray[np.floating] | None,
+    localize: bool,
+    indices: RGIArtifacts | None,
+) -> tuple[npt.NDArray[np.floating], RGIArtifacts]:
+    if interp.method != "linear":
+        msg = "Parameter 'indices' is only supported for 'method=linear'"
+        raise ValueError(msg)
+    if localize:
+        msg = "Parameter 'indices' is only supported for 'localize=False'"
+        raise ValueError(msg)
+
+    if indices is None:
+        assert xi is not None, "xi must be provided if indices is None"
+        out_of_bounds = interp._prepare_xi_simple(xi)
+        xi_indices, norm_distances = interp._find_indices(xi.T)
+        indices = RGIArtifacts(xi_indices, norm_distances, out_of_bounds)
+
+    out = interp._evaluate_linear(indices.xi_indices, indices.norm_distances)
+    out = interp._set_out_of_bounds(out, indices.out_of_bounds)
+    return out, indices
+
+
+@dataclasses.dataclass
+class RGIArtifacts:
+    """An interface to intermediate RGI interpolation artifacts."""
+
+    xi_indices: npt.NDArray[np.int64]
+    norm_distances: npt.NDArray[np.floating]
+    out_of_bounds: npt.NDArray[np.bool_]
+
+
+# ------------------------------------------------------------------------------
+# 1 dimensional interpolation
+# ------------------------------------------------------------------------------
+
+
+class EmissionsProfileInterpolator:
+    """Support for interpolating a profile on a linear or logarithmic scale.
+
+    This class simply wraps :func:`numpy.interp` with fixed values for the
+    ``xp`` and ``fp`` arguments. Unlike :class:`xarray.DataArray` interpolation,
+    the :func:`numpy.interp` automatically clips values outside the range of the
+    ``xp`` array.
+
+    Parameters
+    ----------
+    xp : npt.NDArray[np.floating]
+        Array of x-values. These must be strictly increasing and free from
+        any nan values. Passed to :func:`numpy.interp`.
+    fp : npt.NDArray[np.floating]
+        Array of y-values. Passed to :func:`numpy.interp`.
+    drop_duplicates : bool, optional
+        Whether to drop duplicate values in ``xp``. Defaults to ``True``.
+
+    Examples
+    --------
+    >>> xp = np.array([3, 7, 10, 30], dtype=float)
+    >>> fp = np.array([0.1, 0.2, 0.3, 0.4], dtype=float)
+    >>> epi = EmissionsProfileInterpolator(xp, fp)
+    >>> # Interpolate a single value
+    >>> epi.interp(5)
+    np.float64(0.150000...)
+
+    >>> # Interpolate a single value on a logarithmic scale
+    >>> epi.log_interp(5)
+    np.float64(1.105171...)
+
+    >>> # Demonstrate speed up compared with xarray.DataArray interpolation
+    >>> import time, xarray as xr
+    >>> da = xr.DataArray(fp, dims=["x"], coords={"x": xp})
+
+    >>> inputs = [np.random.uniform(0, 31, size=200) for _ in range(1000)]
+    >>> t0 = time.perf_counter()
+    >>> xr_out = [da.interp(x=x.clip(3, 30)).values for x in inputs]
+    >>> t1 = time.perf_counter()
+    >>> np_out = [epi.interp(x) for x in inputs]
+    >>> t2 = time.perf_counter()
+    >>> assert np.allclose(xr_out, np_out)
+
+    >>> # We see a 100 fold speed up (more like 500x faster, but we don't
+    >>> # want the test to fail!)
+    >>> assert t2 - t1 < (t1 - t0) / 100
+    """
+
+    def __init__(
+        self,
+        xp: npt.NDArray[np.floating],
+        fp: npt.NDArray[np.floating],
+        drop_duplicates: bool = True,
+    ) -> None:
+        if drop_duplicates:
+            # Using np.diff to detect duplicates ... this assumes xp is sorted.
+            # If xp is not sorted, an ValueError will be raised in _validate
+            mask = np.abs(np.diff(xp, prepend=np.inf)) < 1e-15  # small tolerance
+            xp = xp[~mask]
+            fp = fp[~mask]
+
+        self.xp = np.asarray(xp)
+        self.fp = np.asarray(fp)
+        self._validate()
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(xp={self.xp}, fp={self.fp})"
+
+    def _validate(self) -> None:
+        if not len(self.xp):
+            msg = "xp must not be empty"
+            raise ValueError(msg)
+        if len(self.xp) != len(self.fp):
+            msg = "xp and fp must have the same length"
+            raise ValueError(msg)
+        if not np.all(np.diff(self.xp) > 0.0):
+            msg = "xp must be strictly increasing"
+            raise ValueError(msg)
+        if np.any(np.isnan(self.xp)):
+            msg = "xp must not contain nan values"
+            raise ValueError(msg)
+
+    def interp(self, x: npt.NDArray[np.floating]) -> npt.NDArray[np.floating]:
+        """Interpolate x against xp and fp.
+
+        Parameters
+        ----------
+        x : npt.NDArray[np.floating]
+            Array of x-values to interpolate.
+
+        Returns
+        -------
+        npt.NDArray[np.floating]
+            Array of interpolated y-values arising from the x-values. The ``dtype`` of
+            the output array is the same as the ``dtype`` of ``x``.
+        """
+        # Need to explicitly cast back to x.dtype
+        # https://github.com/numpy/numpy/issues/11214
+        x = np.asarray(x)
+        dtype = np.result_type(x, np.float32)
+        return np.interp(x, self.xp, self.fp).astype(dtype)
+
+    def log_interp(self, x: npt.NDArray[np.floating]) -> npt.NDArray[np.floating]:
+        """Interpolate x against xp and fp on a logarithmic scale.
+
+        This method composes the following three functions.
+            1. :func:`numpy.log`
+            2. :meth:`interp`
+            3. :func:`numpy.exp`
+
+        Parameters
+        ----------
+        x : npt.NDArray[np.floating]
+            Array of x-values to interpolate.
+
+        Returns
+        -------
+        npt.NDArray[np.floating]
+            Array of interpolated y-values arising from the x-values.
+        """
+        return np.exp(self.interp(np.log(x)))