interpolars 2.2.0

Interpolation plugin for Polars

interpolars

interpolars is a small Polars plugin that does N-dimensional interpolation from a source "grid" (your DataFrame) onto an explicit target DataFrame, with optional spherical-geometry-aware geospatial interpolation for lat/lon data.

It supports:

• 1D/2D/3D/... multilinear interpolation (interpolate_nd)
• Geospatial interpolation on lat/lon coordinates with IDL wrapping, pole handling, and scattered-data methods (interpolate_geospatial)
• Multiple value columns in one call
• Target passthrough columns (e.g. labels/metadata)
• Grouped interpolation over "extra" coordinate dims (e.g. group by time and interpolate over latitude/longitude for each time slice)
• Non-float coordinate dtypes such as Date and Duration (they are cast internally for interpolation math; group keys preserve dtype in output)
• Configurable NaN/Null handling (handle_missing): error, drop, fill with a constant, or nearest-neighbor fill
• Boundary extrapolation (extrapolate): linearly project beyond the source grid instead of clamping

---

Installation

This repo is built with maturin and managed with uv.

• As a local editable/dev install (recommended for hacking on it):

cd /path/to/interpolars
uv sync --dev

• Run tests:

cd /path/to/interpolars
uv run pytest

Notes:

• Python: >= 3.12 (see pyproject.toml)
• Polars: pinned to polars==1.37.1

---

The API

Two public functions are available:

• interpolate_nd -- general N-dimensional interpolation on rectilinear grids
• interpolate_geospatial -- latitude/longitude interpolation with spherical-geometry awareness (IDL wrapping, pole handling, scattered-data methods)

Both return a Polars expression (pl.Expr).

---

interpolate_nd

from interpolars import interpolate_nd

expr = interpolate_nd(
    expr_cols_or_exprs=["x", "y"],            # source coordinate columns/exprs
    value_cols_or_exprs=["value_a", "value_b"],  # source value columns/exprs
    interp_target=target_df,                  # DataFrame with target coordinates (+ metadata)
    handle_missing="error",                   # "error" | "drop" | "fill" | "nearest"
    fill_value=None,                          # required when handle_missing="fill"
    extrapolate=False,                        # True → linear extrapolation at boundaries
)

You typically use it inside LazyFrame.select:

import polars as pl
from interpolars import interpolate_nd

out = (
    source_df.lazy()
    .select(interpolate_nd(["x", "y"], ["value"], target_df))
    .collect()
)

---

Output shape and how to consume it

interpolate_nd(...) produces a single struct column named "interpolated".

That struct contains, in order:

• all columns from interp_target (including metadata like label)
• any "extra/group" coordinate dims from the source (see next section)
• all interpolated value fields

To "flatten" the result into normal columns, use unnest:

flat = (
    source_df.lazy()
    .select(interpolate_nd(["x", "y"], ["value"], target_df))
    .unnest("interpolated")
    .collect()
)

Or access fields directly:

only_value = (
    source_df.lazy()
    .select(interpolate_nd(["x"], ["value"], target_df).struct.field("value").alias("value"))
    .collect()
)

---

Grouped interpolation over extra coordinate dims (e.g. time slices)

If the source coordinate columns include fields that do not exist in interp_target, those fields are treated as grouping dimensions.

Example:

• source coords: ["latitude", "longitude", "time"]
• target df columns: ["latitude", "longitude", "label"]
• values: ["2m_temp", "precipitation"]

Then time is a group key:

1. The source rows are grouped by unique time
2. Interpolation runs over (latitude, longitude) within each time group
3. Results are concatenated, producing len(target_df) * n_times rows

import polars as pl
from interpolars import interpolate_nd

target = pl.DataFrame(
    {
        "latitude": [0.25, 0.75],
        "longitude": [0.50, 0.25],
        "label": ["a", "b"],
    }
)

out = (
    source_df.lazy()
    .select(
        interpolate_nd(
            ["latitude", "longitude", "time"],
            ["2m_temp", "precipitation"],
            target,
        )
    )
    .unnest("interpolated")
    .collect()
)

Output order is deterministic:

• target rows are repeated per group
• groups are ordered by ascending group key (e.g. ascending time)

---

Date and Duration coordinates

Coordinate columns can be pl.Date, pl.Datetime, and pl.Duration (and other numeric-like dtypes). The plugin will cast coordinates internally for interpolation computations.

Example (Date as an interpolation axis):

from datetime import date
import polars as pl
from interpolars import interpolate_nd

source = pl.DataFrame(
    {
        "d": pl.Series("d", [date(2020, 1, 1), date(2020, 1, 3)], dtype=pl.Date),
        "value": [0.0, 2.0],
    }
)
target = pl.DataFrame(
    {
        "d": pl.Series("d", [date(2020, 1, 2)], dtype=pl.Date),
        "label": ["mid"],
    }
)

out = (
    source.lazy()
    .select(interpolate_nd(["d"], ["value"], target))
    .unnest("interpolated")
    .collect()
)

Example (Duration as an interpolation axis):

import polars as pl
from interpolars import interpolate_nd

source = pl.DataFrame(
    {
        "dt": pl.Series("dt", [0, 10_000], dtype=pl.Duration("ms")),
        "value": [0.0, 10.0],
    }
)
target = pl.DataFrame(
    {
        "dt": pl.Series("dt", [5_000], dtype=pl.Duration("ms")),
        "label": ["half"],
    }
)

out = (
    source.lazy()
    .select(interpolate_nd(["dt"], ["value"], target))
    .unnest("interpolated")
    .collect()
)

---

Handling NaN and Null values (handle_missing)

By default, any NaN or Null in source coordinates or values will raise an error. You can change this with the handle_missing parameter:

Mode	Coords with NaN/Null	Values with NaN/Null
"error" (default)	Error	Error
"drop"	Drop row	Drop row
"fill"	Drop row	Replace with fill_value
"nearest"	Drop row	Replace with nearest valid grid point's value

• Rows with NaN/Null in coordinate columns are always dropped (except in "error" mode, which raises). A grid point with no location cannot be meaningfully filled.
• fill_value is required when handle_missing="fill" and ignored otherwise.
• "nearest" finds the closest valid grid point by Euclidean distance in coordinate space.

# Drop any source rows that have NaN or Null in coords or values
out = (
    source_df.lazy()
    .select(interpolate_nd(["x", "y"], ["value"], target_df, handle_missing="drop"))
    .collect()
)

# Replace NaN/Null values with 0.0 (NaN coords are dropped)
out = (
    source_df.lazy()
    .select(
        interpolate_nd(
            ["x", "y"], ["value"], target_df,
            handle_missing="fill", fill_value=0.0,
        )
    )
    .collect()
)

# Replace NaN/Null values with the nearest valid grid point's value
out = (
    source_df.lazy()
    .select(interpolate_nd(["x", "y"], ["value"], target_df, handle_missing="nearest"))
    .collect()
)

Note: "drop" can cause "missing corner point" errors if the remaining grid is no longer a full cartesian product after removing rows. "fill" and "nearest" preserve the grid structure.

---

Boundary extrapolation (extrapolate)

By default, target points outside the source grid are clamped to the nearest boundary value. Set extrapolate=True to linearly project from the two nearest grid points along each axis instead:

import polars as pl
from interpolars import interpolate_nd

source = pl.DataFrame({"x": [0.0, 1.0, 2.0], "value": [0.0, 10.0, 20.0]})

# x=3.0 is outside [0, 2]; extrapolate from slope of (1,10)→(2,20)
target = pl.DataFrame({"x": [3.0]})

out = (
    source.lazy()
    .select(interpolate_nd(["x"], ["value"], target, extrapolate=True))
    .unnest("interpolated")
    .collect()
)
# value = 30.0  (linear projection)

Without extrapolate=True, the same query would clamp to the boundary and return 20.0.

handle_missing and extrapolate compose freely -- for example, handle_missing="nearest", extrapolate=True fills NaN values with the nearest neighbor and extrapolates at boundaries.

---

interpolate_geospatial

For latitude/longitude data, interpolate_geospatial provides spherical-geometry-aware interpolation with automatic International Date Line wrapping, pole averaging, and support for both gridded and scattered source data.

from interpolars import interpolate_geospatial

expr = interpolate_geospatial(
    source_lat="lat",                      # source latitude column (degrees)
    source_lon="lon",                      # source longitude column (degrees)
    value_cols_or_exprs=["temperature"],   # value column(s) to interpolate
    interp_target=target_df,               # DataFrame with target lat/lon (+ metadata)
    handle_missing="error",                # "error" | "drop" | "fill" | "nearest"
    fill_value=None,                       # required when handle_missing="fill"
    extrapolate=False,                     # True → extrapolate at boundaries
    method="tensor_product",               # "tensor_product" | "slerp" | "idw" | "rbf"
    # keyword-only parameters:
    tensor_method="linear",                # 1-D sub-method for tensor_product
    power=2.0,                             # distance exponent for idw
    k_neighbors=0,                         # nearest neighbors for idw/rbf (0 = all for idw)
    rbf_kernel="thin_plate_spline",        # kernel for rbf
    rbf_epsilon=None,                      # shape param for rbf (None = auto)
    lon_range="auto",                      # "signed_180" | "unsigned_360" | "auto"
)

Usage follows the same pattern as interpolate_nd:

import polars as pl
from interpolars import interpolate_geospatial

out = (
    source_df.lazy()
    .select(interpolate_geospatial("lat", "lon", ["temperature"], target_df))
    .unnest("interpolated")
    .collect()
)

---

Geospatial interpolation methods

Four methods are available, selected via the method parameter:

Method	Input requirement	Description
"tensor_product" (default)	Rectilinear grid	Tensor-product interpolation with longitude wrapping, pole averaging, and ghost points for periodic grids. Supports all 1-D sub-methods via tensor_method.
"slerp"	Rectilinear grid	Bilinear interpolation using SLERP-derived angular fraction weights along parallels. More accurate than standard bilinear near the poles and for large grid cells. Linear only.
"idw"	Any (including scattered)	Inverse Distance Weighting using Haversine (great-circle) distance. Tune via power and k_neighbors.
"rbf"	Any (including scattered)	Local Radial Basis Function interpolation using Haversine distance. Solves a k x k linear system per target point. Tune via rbf_kernel, rbf_epsilon, and k_neighbors.

Tensor-product sub-methods (tensor_method)

When method="tensor_product", the tensor_method parameter selects the 1-D interpolation method applied along each axis:

"linear" (default), "nearest", "cubic", "pchip", "akima", "makima"

IDW tuning

• power (default 2.0): distance exponent. Higher values give more weight to nearby points.
• k_neighbors (default 0): number of nearest source points to use. 0 means use all.

RBF tuning

• rbf_kernel: "linear", "thin_plate_spline" (default), "cubic", "gaussian", "multiquadric", "inverse_multiquadric"
• rbf_epsilon: shape parameter (None = auto-detect from median pairwise distance)
• k_neighbors: number of nearest neighbors for the local solve (default 20 for RBF)

---

Longitude convention (lon_range)

The lon_range parameter controls how longitude values are normalized:

Mode	Range	When to use
"signed_180"	[-180, 180)	Source data uses negative longitudes for the Western hemisphere
"unsigned_360"	[0, 360)	Source data uses 0-360 convention
"auto" (default)	Detected from source	Uses signed_180 if any source longitude is negative, otherwise unsigned_360

# Explicit signed_180: source data with 350° is normalized to -10°
result = (
    source_df.lazy()
    .select(
        interpolate_geospatial(
            "lat", "lon", ["v"], target_df,
            lon_range="signed_180",
        )
    )
    .unnest("interpolated")
    .collect()
)

---

Important constraints / behavior

• NaN/Null handling is configurable via handle_missing (see above). The default ("error") raises on any NaN or Null.
• Grid requirement: interpolate_nd, tensor_product, and slerp require a full cartesian grid (per group) -- every corner must exist. idw and rbf work on arbitrary scattered points.
• Out-of-bounds targets: clamped by default; set extrapolate=True for linear extrapolation.
• Duplicate names: value field names cannot collide with interp_target columns (and group fields cannot collide either); collisions error.

---

Project layout

• src/interpolars/__init__.py: Python API wrapper (interpolate_nd + interpolate_geospatial)
• src/expressions.rs: the Polars expression implementation (Rust)
• tests/: pytest suite with examples (including grouped, Date/Duration, and geospatial coverage)