petekTools

Scattered-data gridding & geostatistics kernels for Rust — the numerics layer the ecosystem is missing.

Rust has excellent crates for linear algebra (faer), statistics and distributions (statrs, rand_distr), FFT (rustfft), and spatial indexing (kiddo, rstar). What it has lacked is a production-grade way to turn scattered (x, y, z) observations into a regular grid — minimum-curvature surfaces, inverse-distance weighting, nearest-neighbour fills. petekTools fills exactly that gap, and curates the rest behind one small front-door.

If you have points and need a surface — a depth grid from well picks, a property map from samples, any scattered field on a regular lattice — this is the crate.

Why reach for it

• Real gridding methods, not toys. Briggs minimum-curvature (biharmonic SOR), inverse-distance weighting, and nearest-neighbour — the workhorses, with their defaults and tolerances stated and tested, not hand-waved.
• Warm-start / incremental re-gridding. Editing a surface point-by-point? Re-solving from scratch each time is wasteful. Seed the solver from the prior field and it converges in a fraction of the iterations — ~4–7× faster on a typical structural edit, rising to ~114× in the near-converged incremental limit (measured: 1.55 ms → 13.6 µs, 64 points → 40×40), converging to the same field. A stateful [ConvergentGridder] makes interactive, one-control-at-a-time refinement cheap.
• Type-agnostic by design. Kernels speak a plain [Lattice] + [[f64; 3]] rows and return ndarray::Array2<f64> — never a domain type. Any regular, rotatable areal lattice (the IRAP/RMS model) maps on field-for-field, so adoption is a conversion at the call site, not a rewrite.
• Deterministic and honest. No RNG, no silent clamping; named, cited constants. Analytic cases are asserted as tests — a linear trend is the exact minimum-curvature solution, IDW is exact at coincident samples.
• A pure leaf. Depends only on general-purpose numeric crates. No I/O, no domain model, no heavy framework to adopt — drop it in.
• Binding-friendly. Owned inputs, no public lifetimes on kernels; PyO3 bindings are a planned thin layer over this same surface.

Install

cargo add petektools

Quick start

use petektools::{grid, GridMethod, Lattice};

// A 100×80 grid, 25 m spacing, origin at (1000, 2000).
let lattice = Lattice::regular(1000.0, 2000.0, 25.0, 25.0, 100, 80);

// Scattered observations: [x, y, z] rows.
let points = [
    [1010.0, 2008.0, 12.5],
    [1240.0, 2300.0, 18.1],
    [1880.0, 3100.0,  9.4],
];

// Interpolate onto the grid → an (ncol × nrow) Array2<f64>;
// undefined nodes are NaN.
let surface = grid(&points, &lattice, GridMethod::MinimumCurvature).unwrap();

Methods

Warm-start an incremental re-grid

use petektools::{grid, grid_min_curvature_seeded, GridMethod, Lattice};

let lattice = Lattice::regular(0.0, 0.0, 1.0, 1.0, 64, 64);
let points = [[1.0, 1.0, 10.0], [60.0, 60.0, 40.0]];

let cold = grid(&points, &lattice, GridMethod::MinimumCurvature).unwrap();

// Later, after nudging the data: relax from the prior field instead of
// cold-starting. A None / wrong-shape seed falls back to a cold solve.
let warm = grid_min_curvature_seeded(&points, &lattice, Some(&cold)).unwrap();

Interactive refinement with ConvergentGridder

use petektools::{ConvergentGridder, Lattice};

let lattice = Lattice::regular(0.0, 0.0, 1.0, 1.0, 64, 64);
let points = [[1.0, 1.0, 10.0], [60.0, 60.0, 40.0]];

let mut gridder = ConvergentGridder::new(&points, &lattice).unwrap();

// Pin node (32, 40) to 25.0 as a hard constraint and re-solve incrementally.
let field = gridder.add_control(32, 40, 25.0);   // warm — cheap
// … add more controls; each call returns the updated field.

Geometry: Lattice

A regular, rotatable areal lattice (the IRAP/RMS model): origin, spacing, node counts, CCW rotation, and an optional y-flip. It carries the forward map node_xy(i, j) and its inverse xy_to_ij(x, y), plus a bbox() — everything a kernel needs to place a sample on the grid.

Design

• One job. Scattered-data gridding / geostatistics — the gap. Everything else (linear algebra, stats, neighbour search) is curated from mature crates, never reimplemented.
• Type-agnostic kernels. Lattice + [[f64; 3]] in, ndarray out. No domain types, no I/O.
• Numerical honesty. Deterministic, documented to a stated tolerance, with analytic tests as the safety net.

See SPEC.md for the design constitution and API.md for the locked public contract.

Status & roadmap

The public contract — Lattice, GridMethod, grid, the warm-start entries — is locked and analytically tested. Also shipped: ordinary kriging behind a Gridder trait, the curated stats / sampling front-doors over statrs / rand_distr, and the units / container modules. On the roadmap: a PyO3 wheel (and RBF backends if a need appears).

License

Apache-2.0 — see LICENSE and NOTICE.