scientific-computing-system

A pure-Python computational science platform for numerical methods, modeling, validation, uncertainty, scientific workflows, dimensional analysis, and reproducible research.

Docs · Getting Started · API Reference · Cookbook · Benchmarks · Project · Changelog · Contributing · Security · Code of Conduct · Citation · Paper (JOSS) · Discussions · Issues

I wrote this because NumPy and SciPy are incredible, but they're also 20 years old and carry two decades of design decisions that don't always make sense anymore.

Download counts: PyPI stats

This is a from-scratch rethinking of what scientific computing in Python could look like if we started today. No C extensions, no Fortran legacy, no dependency hell. Just Python, type hints, and algorithms that are actually readable.

The import name is cds (the distribution is scientific-computing-system). The zero-dependency core needs nothing but the standard library. The sibling NumPy build of this work is scientific-computing-system-2.0; related projects are listed in docs/related.md.

Install

Both channels are published at the same version. They are not equally verifiable:

• PyPI — carries a PEP 740 provenance attestation (Sigstore-signed, binding the file digest to the publishing workflow). Verified on 2.2.1.
• npm — carries no attestation; the npmjs.com trusted publisher is not registered yet, so its publish path cannot mint one. npm does expose dist.integrity (sha512), which is a content pin, not proof of origin, and dist.signatures, which is npm's registry transport signature, not build provenance.

Prefer PyPI for anything security-sensitive. SECURITY.md has the per-channel table, the published digests to pin, the commands to verify them, and what to register to close the npm gap.

What's inside

src/cds/ holds 34 subpackages plus causal.py and sensitivity.py. That is 35 public feature modules — the cds CLI entry point is the 36th subpackage — exporting 505 names (342 functions, 163 classes), all pure standard library. tests/test_readme_surface.py derives these numbers from the package, so they cannot drift silently.

The authoritative per-module list is the API reference.

!!! note "cds modules is a curated subset" cds modules prints 26 entries. It is not the full catalog: it omits bayes, causal, core, fractals, genetics, infotheory, interpolate, pde, and wavelets. Use the API reference or src/cds/ for the complete set.

Runnable examples and the dashboard

• examples/ — 35 runnable scripts and notebooks (33 .py, 2 .ipynb), each self-contained and dependency-free unless noted. Every tutorial page in the docs has a matching example.
• dashboard/app.py — a Streamlit dashboard: cds dashboard, or pip install "scientific-computing-system[dashboard]".

Quick Start

pip install scientific-computing-system

from cds.math_utils import svd
from cds.diffeq import solve_system
from cds.stats import linear_regression

# 1. SVD — result is an SVDResult, not a tuple
result = svd([[1.0, 2.0], [3.0, 4.0]])
print("singular_values =", [round(v, 6) for v in result.singular_values])


# 2. Damped harmonic oscillator: y'' = -y - 0.1 y'
def rhs(t, y):
    return [y[1], -y[0] - 0.1 * y[1]]


t, y = solve_system(rhs, 0.0, [1.0, 0.0], 50.0, dt=0.01)
print("final t =", round(t[-1], 2), "| final y =", [round(v, 8) for v in y[-1]])

# 3. Least-squares line fit
fit = linear_regression([1.0, 2.0, 3.0, 4.0, 5.0], [2.1, 3.9, 6.2, 7.8, 10.1])
print(f"slope = {fit.slope:.4f} | intercept = {fit.intercept:.4f} | r^2 = {fit.r_squared:.4f}")

Real output (executed verbatim; pasted from an actual run):

singular_values = [5.464986, 0.365966]
final t = 50.0 | final y = [0.07638443, 0.0264785]
slope = 1.9900 | intercept = 0.0500 | r^2 = 0.9973

CLI

The package installs a cds command with 13 subcommands:

$ cds --version
System version 2.2.2
$ cds constants        # physical constants table

Optional extras

The core is zero-dependency. Everything below is opt-in:

pip install "scientific-computing-system[scientific]"  # NumPy, SciPy, statsmodels, scikit-learn, SymPy, Z3
pip install "scientific-computing-system[io]"          # HDF5 + NetCDF
pip install "scientific-computing-system[plot]"        # matplotlib
pip install "scientific-computing-system[pandas]"      # DataFrame bridge: cds.data_analysis.pandas_io
pip install "scientific-computing-system[dashboard]"   # Streamlit dashboard

The catch

Pure Python is slower than NumPy on dense numerics: the committed benchmark artifact records 100×100 matrix multiplication at 0.0696 s against NumPy's 0.000060 s, about 1155× slower. Those numbers are from one CI run whose platform, CPU, and library versions were not recorded, so treat them as order-of-magnitude only — docs/benchmarks.md states exactly what that artifact does and does not support, and docs/why-pure-python.md covers when to reach for NumPy or the 2.0 build instead.

Contributing

pip install -e ".[dev]"
pytest            # CI also enforces 100% blended coverage (statement + branch)
mkdocs serve      # docs at http://127.0.0.1:8000/

Issues and PRs welcome; see CONTRIBUTING.md and CODE_OF_CONDUCT.md.

License

MIT — see LICENSE.

Cite via CITATION.cff. codemeta.json carries the same metadata in codemeta form, and paper.md / paper.bib are the JOSS manuscript.