carnopy 0.1.0a3

CLI-first thermophysical dataset generation and leakage-aware ML preparation from thermodynamic backends, with an optional Linux-first desktop GUI.

Carnopy

CLI-first thermophysical dataset generation and leakage-aware ML preparation from thermodynamic backends, with an optional Linux-first desktop GUI.

Alpha software: public interfaces and generated schemas may still change before the stable 0.1.0 release.

Carnopy is not a thermodynamic property model. It orchestrates configured property backends, validates deterministic sampling, preserves failed states as diagnostics, and emits stable tabular data with provenance. Generated values are synthetic backend output, not experimental data or backend-independent ground truth.

Milestone 1 supports pure fluids through CoolProp and three modes:

• property_table: temperature-pressure state tables;
• saturation_table: saturated-liquid and saturated-vapor endpoint rows;
• vapor_mass_fraction_table: two-phase states over vapor mass fraction.

Carnopy has two user interfaces:

• Command-line interface: version 0.1.0a3 provides the complete dataset, sweep, inspection, plotting, and preparation workflow.
• Desktop GUI: version 0.1.0a3 provides an optional Linux-first PySide6 application for dataset configuration, generation, inspection, and plotting.

Installation

Version 0.1.0a3

Install the CLI-first base package:

python -m pip install "carnopy==0.1.0a3"

Add Matplotlib plotting support:

python -m pip install "carnopy[viz]==0.1.0a3"

Add SafeTensors preparation exports:

python -m pip install "carnopy[ml]==0.1.0a3"

Add optional scikit-learn preparation diagnostics:

python -m pip install "carnopy[analysis]==0.1.0a3"

Add the desktop application:

python -m pip install "carnopy[app]==0.1.0a3"

Install every optional capability:

python -m pip install "carnopy[all]==0.1.0a3"

For an isolated CLI installation:

uv tool install "carnopy==0.1.0a3"

To include every optional capability:

uv tool install "carnopy[all]==0.1.0a3"

The 0.1.0a3 base install remains CLI-first and does not install Matplotlib, SafeTensors, scikit-learn, or PySide6. Its lightweight carnopy-app --help, carnopy-gui --help, and version entry points remain available, while opening the desktop requires the app extra. The app extra installs PySide6 Essentials and Matplotlib. The all extra is the dependency union of viz, ml, analysis, and app. The desktop launchers remain separate from the CLI as carnopy-app and carnopy-gui. PyArrow remains a core dependency because Parquet is a first-class output format.

Development checkout

Run the desktop directly from a source checkout:

git clone https://github.com/gcalpay/carnopy.git
cd carnopy
uv sync --locked --extra app --group dev
uv run --locked carnopy-gui

The source-tree app extra has the same optional dependency boundary.

Quick start

carnopy init property_table my-dataset.yaml
# Edit the generated YAML, then:
carnopy generate my-dataset.yaml
carnopy inspect outputs/<run>
carnopy plot outputs/<run> \
  --kind property-curves \
  --property mass_density \
  --x temperature

The normal workflow is:

init → edit → optional validate → generate/sweep → inspect → optional plot → optional prepare

For repository development:

uv sync --locked --extra all --group dev
uv run --locked carnopy --help

The current desktop development build can be opened from a source checkout:

uv run --locked carnopy-gui

To preselect a workspace without initializing it silently:

uv run --locked carnopy-gui --workspace /path/to/workspace

Qt normally selects its platform integration automatically. If WSLg leaves drop-down popups visible after they close, and the system Qt XCB runtime is available, start the application explicitly with XCB:

uv run --locked carnopy-gui --qt-platform xcb --workspace /path/to/workspace

The corresponding explicit Wayland selection is --qt-platform wayland; omit the option or use auto for normal Qt platform detection.

GUI-1 currently provides workspace lifecycle, a worker-validated editor for all three dataset configuration modes, dataset validation and generation, output inspection, bounded table previews, job diagnostics, guarded staging recovery, and manual plot rendering with immediate confirmed force-stop. Plot rendering uses Matplotlib only in a short-lived worker and writes a no-overwrite-safe image plus .plot.json sidecar. PNG and SVG exports preview automatically in Qt with fit, zoom, 100%, and panning controls. PDF exports open only after an explicit user action.

Guide

• Workflow details
• Configuration
• Properties
• Visualization
• Generated outputs and provenance
• Python API
• Desktop GUI
• ML preparation roadmap
• Architecture map
• Scientific limitations
• Development and contribution
• Project status and roadmap

Workflow details

init → edit → optional validate → generate/sweep → inspect → optional plot → optional prepare

Create a starter configuration:

carnopy init property_table my-dataset.yaml

init reads the selected template packaged inside the installed carnopy module and writes a new file at the path you provide. For example, when the current directory is /home/cfd/carnopy/:

carnopy init property_table property.yaml

creates:

/home/cfd/carnopy/property.yaml

from the packaged property_table.yaml template. It does not modify or move the packaged source.

Use --full to append the exhaustive commented reference for current samplers, properties, units, output formats, visualization kinds, filters, series selection, scales, and identity rules:

carnopy init property_table my-dataset.yaml --full

The active mode-specific configuration remains valid. Normal templates remain concise. Both forms refuse to overwrite an existing property.yaml. A relative output path is resolved from the current working directory; an absolute path is written exactly where specified.

Available dataset modes:

property_table
saturation_table
vapor_mass_fraction_table

Additional workflow/template types:

model_sweep

model_sweep is not a dataset mode. It creates a sweep bundle containing one immutable child dataset run per selected model plus comparison artifacts.

Discover backend fluids and semantic properties:

carnopy fluids                 # HEOS default
carnopy fluids --model pr      # model-specific availability
carnopy properties

Edit the YAML, optionally validate it, then generate an immutable run:

carnopy validate my-dataset.yaml
carnopy generate my-dataset.yaml

generate validates automatically. The separate validate command is useful for scripts and early feedback, but does not evaluate thermodynamic rows.

After generation, inspect the run before choosing a plot:

carnopy inspect outputs/<run>

The inspection lists fluids, sampling levels, emitted properties, compatible plot kinds, valid/invalid rows, phase and failure counts, property ranges, available curve-series fields, supported display units, and copyable commands.

Use structured output in scripts or create a visualization-only starter file for the immutable run:

carnopy inspect outputs/<run> --format json
carnopy inspect outputs/<run> --write-visualization plots.yaml
carnopy plot outputs/<run> --config plots.yaml

The writer uses exclusive creation and refuses to replace an existing YAML file. It does not evaluate thermodynamic states or create a figure.

To choose a different output root:

carnopy generate \
  configs/cyclopentane_vapor_fraction_pressure.yaml \
  --out outputs/manual-test

The run is created directly under that root. Copy the exact path printed after Output directory:; do not prepend the output root again:

# Example only; replace this with the exact path printed by your run.
RUN_DIR="outputs/manual-test/20260621T172006Z_vapor_fraction_c8e28e9f"

Run names use UTC creation time, a short mode label, and the first eight hexadecimal characters of the unique run_id. Full identities and hashes remain in metadata.json.

Use command-specific help for the complete current interface:

carnopy --help
carnopy generate --help
carnopy plot --help

Configuration

Schema version 2 requires an explicit dataset document type and CoolProp thermodynamic model:

schema_version: 2
document_type: dataset
backend:
  name: coolprop
  model: heos
mode: property_table
fluids: [Propane]

grid:
  temperature:
    kind: linspace
    start: 20
    stop: 100
    num: 5
    unit: degC
  pressure:
    kind: linspace
    start: 1
    stop: 20
    num: 5
    unit: bar

properties:
  - specific_enthalpy
  - mass_density

outputs:
  # Omit this section to keep the same default.
  dataset_formats: [csv, parquet]

Schema version 1 configuration files are intentionally rejected with a concise migration message. Existing generated run directories remain readable.

CoolProp models

Supported model names:

Model	Meaning	Current capability notes
heos	CoolProp Helmholtz-energy equations and associated ancillary/transport models	Supports the full current Carnopy property registry, subject to fluid/state limitations.
pr	Peng-Robinson cubic equation of state	No viscosity, thermal conductivity, Prandtl number, surface tension, or usable triple-point temperature.
srk	Soave-Redlich-Kwong cubic equation of state	No viscosity, thermal conductivity, Prandtl number, surface tension, or usable triple-point temperature.

Model selection is part of the executable scientific specification and changes spec_id. The selected model is recorded in every generated row, metadata, reports, and normalized configuration. HEOS is the starter default, not experimental truth. PR and SRK are alternative model assumptions, not accuracy rankings.

Reference-dependent enthalpy, entropy, and internal energy can differ between models even after each model-qualified fluid is reset to CoolProp DEF. Absolute values must not be compared across model/reference conventions without an explicit scientific basis.

Model sweeps

Model sweeps compare emitted values from several CoolProp models without performing extra thermodynamic evaluations during comparison:

carnopy init model_sweep sweep.yaml
carnopy sweep sweep.yaml

The sweep document type is separate from dataset generation:

schema_version: 2
document_type: model_sweep
backend:
  name: coolprop
  models: [heos, pr, srk]
  reference_model: heos
mode: property_table
fluids: [Propane]
grid:
  temperature: {kind: linspace, start: 280, stop: 340, num: 5, unit: K}
  pressure: {kind: linspace, start: 1, stop: 5, num: 5, unit: bar}
properties: [mass_density]

Each selected model creates a normal immutable child run under the sweep bundle. Comparison artifacts are written as tidy Parquet tables:

comparison/values.parquet
comparison/deltas.parquet

State alignment uses deterministic keys derived from normalized sample indices, not backend-computed floating-point saturation coordinates. The selected reference model is a comparison baseline, not experimental truth. Reference-dependent properties such as enthalpy, entropy, and internal energy are excluded from delta metrics.

Optional sweep-level comparison plots are explicit and separate from child-run visualization. They require the optional visualization dependencies, installed with carnopy[viz] or carnopy[all]. The concise carnopy init model_sweep starter keeps this block commented so no-plot sweeps run in a base installation:

comparison_plots:
  format: png
  plots:
    - name: propane_density_temperature_by_pressure
      kind: property_comparison
      fluid: Propane
      property: mass_density
      x: temperature
      group_by: pressure
      models: [heos, pr, srk]
    - name: propane_density_relative_delta
      kind: property_delta
      fluid: Propane
      property: mass_density
      x: temperature
      group_by: pressure
      models: [pr, srk]
      delta_metric: signed_relative_difference

Stage 4 comparison plots are one-fluid, one-property, one-x-axis side-by-side model value comparisons or model-vs-reference delta plots. Multiple fluids require multiple plot entries.

ML preparation foundation

Preparation is the current ML-pipeline bridge. It reads an existing immutable run or model-sweep bundle and writes deterministic Parquet outputs without calling a thermodynamic backend. Omit scenarios: for a single unsplit prepared table:

carnopy init preparation preparation.yaml
carnopy prepare outputs/<run> --config preparation.yaml --out prepared

Preparation configuration uses its own independent schema version:

schema_version: 1
document_type: preparation
features:
  numeric: [temperature, pressure, mass_density]
  derived: [specific_volume]
categorical_features:
  - field: phase
    encoding: one_hot
    categories: observed
targets: [specific_enthalpy]
auxiliary: [fluid, backend_model, phase, run_id, case_id]
outputs:
  formats: [parquet]

Prepared bundles contain manifest.json, diagnostics.json, dataset_card.md, data/table.parquet, data/provenance.parquet, data/diagnostics.parquet, data/exclusions.parquet, quality_report.json, and data/quality_flags.parquet. table.parquet is the user-facing feature/target table. Provenance and source diagnostics are separated and join back to the table through prepared_row_id. Quality diagnostics are advisory: they summarize robust distributions, finite-value coverage, target ranges, duplicate state candidates, scenario partitions, and exact eligible property-table grid coverage where that shape can be inferred safely. Quality flags stay in a separate long-form table joined through prepared_row_id; they do not exclude rows or change table.parquet. If no source rows can produce the requested representation, Carnopy writes a clearly marked no_eligible_rows bundle without data/table.parquet.

If selected features or targets include reference-dependent properties (specific_enthalpy, specific_entropy, or specific_internal_energy), preparation records the source reference-state context and requires one compatible reference_state_policy/backend/model context across the selected source rows. This allows ordinary same-context training data and rejects misleading mixtures of absolute h, s, or u values from incompatible contexts.

Optional leakage-aware scenarios add deterministic partition artifacts and plain-JSON transformation parameters. Current numeric transformations are log10, standard, minmax, and train-fitted median/IQR robust scaling:

scenarios:
  - name: shuffle_baseline
    kind: shuffle
    seed: 42
    partitions:
      train: 0.8
      validation: 0.1
      test: 0.1
    transformations:
      - field: pressure
        methods: [log10, robust]

  - name: phase_temperature_strata
    kind: stratified_hash
    seed: 42
    partitions:
      train: 0.8
      validation: 0.1
      test: 0.1
    strata:
      categorical: [phase]
      numeric_bins:
        reduced_temperature: [0.8, 0.95, 1.05]

  - name: leave_fluid_out
    kind: leave_fluid_out
    holdouts:
      test: [Isopentane]
    remainder: train

Supported scenarios are unsplit, shuffle, stratified_hash, coordinate_block, range_holdout, leave_fluid_out, phase_holdout, and model_holdout. Shuffle and stratified scenarios keep an exact thermodynamic-state hash in one partition. Numeric strata use only user-declared boundaries; empty or undersized strata fail clearly rather than being silently rebalanced.

Optional matrix and baseline diagnostics are configured separately:

quality:
  matrix_diagnostics:
    correlation_threshold: 0.995
    near_constant_relative_spread: 1.0e-12
  baseline_diagnostics:
    models: [dummy_mean, ridge, hist_gradient_boosting]
    random_seed: 42
    ridge_alpha: 1.0
    histogram_max_iterations: 100

Matrix diagnostics use configured features only and report constants, near-constants, correlations, singular values, numerical rank, effective rank, and conditioning. They fit on each scenario's train partition, or state explicitly that an unsplit report uses all. Baseline diagnostics require carnopy[analysis]; they fit disposable scikit-learn estimators on train and report validation/test MAE, RMSE, and R² per target. Carnopy never persists the estimators or predictions and does not tune, register, deploy, or use them to change prepared rows.

Parquet remains the canonical prepared table. Optional NumPy and SafeTensors exports are derived ML-consumption files:

outputs:
  parquet: true
  arrays:
    formats: [npy, npz, safetensors]
    dtype: float32
    include_auxiliary: false

Array exports require carnopy[ml] or carnopy[all] when SafeTensors is requested. Carnopy records feature/target order, units, shapes, dtype, file hashes, and float32 conversion-error summaries in the manifest. It does not depend on PyTorch or export .pt/.pth files in this release line. Implemented behavior and reviewed future directions are separated in the ML preparation roadmap. Automatic PCA feature replacement, active learning, Manim animations, PyMC workflows, SINDy, optimization, ORC/TFC work, mixtures, new backends, and a professional 3D GUI remain separate reviewed milestones rather than implicit parts of prepare.

Modes

property_table requires temperature and pressure and generates their Cartesian product for every selected fluid.

saturation_table requires exactly one of temperature or pressure. It computes the missing saturation coordinate and emits separate saturated-liquid and saturated-vapor rows.

vapor_mass_fraction_table requires vapor mass fraction plus exactly one of temperature or pressure. Vapor mass fraction is vapor mass divided by total vapor-plus-liquid mass. Carnopy denotes it by $x_{\mathrm{vap}}$ in figures and scientific equations while keeping the explicit public field name vapor_mass_fraction. CoolProp's Q name remains internal to the adapter.

For a pure fluid at fixed saturation temperature or pressure:

• $x_{\mathrm{vap}}=0$ is the saturated-liquid boundary;
• $x_{\mathrm{vap}}=1$ is the saturated-vapor boundary;
• $0<x_{\mathrm{vap}}<1$ is an equilibrium two-phase mixture state.

The endpoint states have definite backend properties. Near-endpoint values such as 0.01 and 0.99 are interior mixture states; they supplement rather than replace the boundaries. For specific enthalpy and specific volume:

h(x_{\mathrm{vap}})
=(1-x_{\mathrm{vap}})h_f+x_{\mathrm{vap}}h_g

\frac{1}{\rho(x_{\mathrm{vap}})}
=\frac{1-x_{\mathrm{vap}}}{\rho_f}
+\frac{x_{\mathrm{vap}}}{\rho_g}

See the CoolProp high-level saturation documentation for the backend definition of the endpoint states.

Samplers

Sampler	Parameters	Behavior
explicit	values	Preserves declared order; values must be finite and unique after SI conversion.
linspace	start, stop, num	Includes both endpoints; supports ascending and descending ranges.
stepspace	start, stop, step	Includes both endpoints; the endpoint must be reachable.
geomspace	start, stop, num	Positive physical endpoints; supports either direction.
logspace	start_exp, stop_exp, num, optional base	Samples exponent space; base must exceed one.

Equal sampler bounds are rejected; use explicit for one value. Geometric and logarithmic sampling is not supported for offset Celsius values or vapor mass fraction. Use Kelvin for geometric temperature grids.

linspace uses uniform increments. For example, start: 1, stop: 5, and num: 5 produce 1, 2, 3, 4, 5. geomspace uses uniform ratios and produces approximately 1, 1.495, 2.236, 3.344, 5 for the same bounds.

Dataset formats

Select generated table formats independently of the scientific specification:

outputs:
  dataset_formats: [csv]

Supported values are csv and parquet. At least one is required. Omitting outputs preserves the default [csv, parquet]. Format selection changes the artifact-generation context and output_request_id, but not spec_id or config.normalized.json.

Units

Supported input units:

temperature: K, degC
pressure: Pa, kPa, MPa, bar
vapor_mass_fraction: "1"

All backend calls and generated numeric columns use SI. Original units and sampler definitions remain recorded in metadata.

Validation rejects non-finite values, non-positive pressure, temperatures at or below absolute zero, vapor mass fractions outside [0, 1], incompatible units, duplicate canonical fluids, and projected runs above 1,000,000 rows.

Validation proves that a configuration is structurally executable. It does not promise that every fluid, state, phase, and requested property will be valid.

Properties

Use carnopy properties for the authoritative installed registry and its HEOS/PR/SRK support columns. Properties globally unsupported by a selected model fail configuration validation before row generation.

Semantic name	Dataset column	Classification
specific_enthalpy	specific_enthalpy_J_kg	backend-provided, reference-dependent
specific_entropy	specific_entropy_J_kgK	backend-provided, reference-dependent
specific_internal_energy	specific_internal_energy_J_kg	backend-provided, reference-dependent
mass_density	mass_density_kg_m3	backend-provided
isobaric_specific_heat_capacity	isobaric_specific_heat_capacity_J_kgK	backend-provided
isochoric_specific_heat_capacity	isochoric_specific_heat_capacity_J_kgK	backend-provided
dynamic_viscosity	dynamic_viscosity_Pa_s	backend-provided
kinematic_viscosity	kinematic_viscosity_m2_s	derived from viscosity and density
thermal_conductivity	thermal_conductivity_W_mK	backend-provided
prandtl_number	prandtl_number	backend-provided
speed_of_sound	speed_of_sound_m_s	backend-provided
molar_mass	molar_mass_kg_mol	fluid constant
critical_temperature	critical_temperature_K	fluid constant
critical_pressure	critical_pressure_Pa	fluid constant
triple_point_temperature	triple_point_temperature_K	fluid constant
surface_tension	surface_tension_N_m	mode/region limited

Derived dependencies may be evaluated internally without being emitted unless explicitly requested. Fluid constants may be repeated in rows and are also summarized in metadata.

Milestone 1 uses strict row validity: failure of any required coordinate, phase, or requested property makes the row invalid. Successfully evaluated values may remain populated while failed values remain null. Requesting a mode-limited property such as surface_tension over a broad state grid can therefore invalidate otherwise usable rows.

Visualization

Visualization is a reproducible view of emitted dataset columns:

• it never calls CoolProp or another thermodynamic backend;
• it never smooths, interpolates, extrapolates, or invents states;
• it preserves invalid and missing gaps;
• it retains markers at emitted samples;
• its identity is separate from scientific dataset identity.

Install carnopy[all] or carnopy[viz] before plotting.

Manual plotting

Supported plot kinds:

property-curves
property-heatmap
xy
pv
ts

Property curves use discrete, colorblind-safe series colors and markers. For property_table, choose the x-axis explicitly:

carnopy plot outputs/<property-run> \
  --kind property-curves \
  --property mass_density \
  --x temperature

Carnopy connects adjacent valid emitted samples with straight line segments as visual guides. It does not smooth or evaluate intermediate states. A sparse series advisory is emitted for connected series with five or fewer samples. Generate a denser source grid for finer thermodynamic resolution. Use SVG or PDF for zoom-independent rendering:

carnopy plot outputs/<run> ... --output figures/plot.svg
carnopy plot outputs/<run> ... --output figures/plot.pdf

For vapor_mass_fraction_table, vapor mass fraction is the x-axis and the sampled saturation pressure or temperature defines the series:

carnopy plot "$RUN_DIR" \
  --kind property-curves \
  --property mass_density \
  --value-scale linear \
  --show

Sampled heatmaps use flat, non-interpolated cells and require at least two unique values on each axis:

carnopy plot "$RUN_DIR" \
  --kind property-heatmap \
  --property specific_enthalpy \
  --color-scale linear

saturation_table does not support property heatmaps because it contains only the two endpoint branches.

Generic x-y plots use numeric semantic fields from emitted columns:

carnopy plot outputs/<property-run> \
  --kind xy \
  --x specific_enthalpy \
  --y specific_entropy \
  --group-by pressure

If more than one independent sampling coordinate remains, --group-by must resolve the ambiguity. Carnopy does not apply hidden grouping precedence.

Conventional thermodynamic diagrams are derived only from emitted columns:

carnopy plot outputs/<run-with-density> --kind pv
carnopy plot outputs/<run-with-entropy> --kind ts

The p-v diagram uses:

specific_volume = 1 / mass_density

The T-s diagram uses emitted entropy and temperature and requires recorded reference-state metadata. Neither command fabricates a saturation dome, critical point, or missing branch.

Exact filters use canonical SI values and never select a nearest neighbor:

carnopy plot "$RUN_DIR" \
  --kind property-curves \
  --property mass_density \
  --filter pressure=200000

Repeat --filter to combine filters with logical AND. Current filter fields are temperature, pressure, vapor mass fraction, phase, and saturation endpoint. Repeat --fluid to select multiple fluids; each fluid receives its own facet.

Select specific members of a curve family with repeatable unit-aware --series options. Values for the same field are combined with logical OR:

carnopy plot outputs/<property-run> \
  --kind property-curves \
  --property specific_enthalpy \
  --x temperature \
  --series pressure=1bar \
  --series pressure=3bar \
  --series pressure=5bar \
  --display-unit temperature=degC \
  --display-unit pressure=bar \
  --display-unit specific_enthalpy=kJ/kg

Series selection is exact after conversion to canonical SI; Carnopy never chooses the nearest emitted level. Supported engineering display conversions cover temperature, pressure, enthalpy, internal energy, entropy, and specific heat capacities. Display conversion changes only figure values and labels, not the immutable SI dataset.

SOURCE may be a run directory, CSV, or Parquet file. Run directories prefer Parquet and verify it against metadata.json. Standalone saturation and vapor-quality files may require --saturation-coordinate pressure or --saturation-coordinate temperature.

Every export writes an image plus .plot.json provenance sidecar under figures/ by default. Existing image or sidecar paths are refused. Finalization uses exclusive same-filesystem hard links: it is no-overwrite-safe, but the two-file pair is not fully crash-atomic.

Configured visualization

An optional top-level visualization section generates figures after the immutable dataset run is finalized:

visualization:
  format: png
  fluids: [Propane]
  display_units:
    pressure: bar

  plots:
    - name: density-vs-temperature
      kind: property_curves
      property: mass_density
      x: temperature
      series:
        pressure: [1bar, 3bar, 5bar]
      display_units:
        temperature: degC
      value_scale: linear

    - name: density-map
      kind: property_heatmap
      property: mass_density
      color_scale: log

    - name: enthalpy-entropy
      kind: xy
      x: specific_enthalpy
      y: specific_entropy
      group_by: pressure

    - name: pressure-specific-volume
      kind: pv

    - name: temperature-entropy
      kind: ts

Supported formats are png, pdf, and svg. Per-plot format and fluids replace their shared values; scales are selected per plot. Per-plot filters are AND-merged with shared filters, and conflicting values for the same field are rejected. Plot names must be unique safe filename slugs. Output paths and interactive display are intentionally not stored in YAML.

Shared or per-plot exact filters use YAML mappings:

visualization:
  filters:
    phase: gas
  plots:
    - name: gas-density
      kind: property_curves
      property: mass_density
      x: temperature
      filters:
        pressure: 100000

Generate with the default figure root:

carnopy generate my-dataset.yaml

Or select another figure root:

carnopy generate my-dataset.yaml \
  --out outputs/manual-test \
  --figures-out figures/manual-test

Configured figures are written to:

<figures-root>/<run-directory-name>/
├── <plot-name>.<format>
├── <plot-name>.plot.json
└── visualization-report.json

The same YAML requests can be applied later to an existing immutable run. The file may be a full Carnopy configuration or a small file containing only a top-level visualization: section:

carnopy plot outputs/<run> \
  --config plots.yaml \
  --figures-out figures

Batch plotting accepts run directories, not standalone CSV/Parquet files. Scientific generation fields in a full config are ignored; requests are validated against the actual emitted run columns. Manual plot options cannot be combined with --config.

Plots execute independently after dataset finalization. A failed plot preserves the immutable run and any successful figures, records outcomes in the report, and makes the CLI exit with code 1. A zero-valid-row dataset retains exit code 3 and records configured plots as skipped.

Visualization settings do not change config.normalized.json, spec_id, or generation_context_id. They receive their own visualization_request_id = viz-<sha256>. Exact YAML bytes still affect the raw configuration hash.

Generated outputs and provenance

outputs/ is a local generated-data directory and is intentionally ignored by Git. Carnopy creates output roots when requested; the repository does not track an empty placeholder or generated example runs.

Each immutable run contains the selected dataset files plus mandatory provenance outputs:

outputs/<run>/
├── dataset.csv          # when requested
├── dataset.parquet      # when requested
├── config.original.yaml
├── config.normalized.json
├── config.reference.yaml # full mode-specific commented configuration helper
├── metadata.json
└── report.json

Runs are staged and then finalized atomically as one directory. Existing final or staging paths are never overwritten.

config.reference.yaml comes from the same packaged source as carnopy init MODE OUTPUT --full. It is created only while staging a new run, included in the artifact inventory and hashes, and never added to or overwritten in an existing run.

Identity layers:

• spec_id: canonical executable scientific specification;
• generation_context_id: specification plus software and artifact context;
• output_request_id: canonical dataset serialization request;
• run_id: one UUID4 execution attempt;
• artifact hashes: exact emitted bytes;
• visualization_request_id: normalized visualization request, independent from dataset identity.

Configuration provenance includes SHA-256 hashes of exact source YAML and canonical materialized SI configuration bytes. Metadata records software versions, backend model, model-qualified reference-state targets, canonical fluids and properties, model capabilities, sampling, failure counts, units, fluid constants, and artifact hashes. Carnopy does not store the host source-config path.

Parquet schema metadata includes the dataset schema version and unit mapping. Figures are derived artifacts outside the run and are not added to immutable dataset artifact hashes.

Python API

from carnopy import generate_dataset, load_config, validate_config

loaded = load_config("my-dataset.yaml")
validation = validate_config("my-dataset.yaml")
result = generate_dataset(
    "my-dataset.yaml",
    output_root="outputs",
    figures_root="figures",
)

When configured visualization exists, result.visualization contains its request ID, status, figure directory, report path, and outcome counts. result.dataset_formats and result.output_request_id describe the selected table serialization independently of the scientific spec_id.

Manual plotting:

from carnopy.visualization import (
    plot_property_heatmap,
    plot_thermodynamic_diagram,
    plot_xy,
)

heatmap = plot_property_heatmap(
    "outputs/<run>",
    property_name="mass_density",
)

xy = plot_xy(
    "outputs/<run>",
    x="specific_enthalpy",
    y="specific_entropy",
    group_by="pressure",
)

pv = plot_thermodynamic_diagram("outputs/<run>", kind="pv")

The returned Matplotlib figure represents an image that has already been exported. Modifying it does not update the image or provenance sidecar.

Desktop GUI

The Linux-first desktop frontend is an optional PySide6 Qt Widgets application over the same Carnopy Python pipelines used by the CLI. Widgets do not parse or invoke CLI output. Scientific validation and execution run in a short-lived worker process through a private, versioned JSON Lines protocol.

The prepared 0.1.0a3 source implementation includes:

• explicit workspace creation, initialization, and reopening;
• workspace-local configs/, outputs/, and figures/ directories;
• structured editing for HEOS, PR, and SRK dataset configurations;
• all three dataset modes, current samplers, units, properties, and output formats;
• guided configured-visualization requests for every current plot kind;
• deterministic read-only YAML preview;
• worker validation before Import or Save;
• exclusive Save As, atomic normal Save, dirty-state prompts, and external-file modification detection;
• optional explicit validation and independent generation of the exact saved configuration, with phases, row progress, cooperative cancellation, and an informational equivalent CLI command;
• direct-child workspace source discovery plus read-only external browsing;
• structured inspection of dataset runs, standalone CSV/Parquet files, model-sweep bundles, and preparation bundles;
• order-preserving table previews fetched in bounded 500-row worker blocks and presented as local 100-row pages;
• workspace-local Validate/Generate job diagnostics and confirmed, guarded cleanup of recognized stale staging directories;
• inspection-driven, session-only plot-request editing for dataset sources;
• a private worker plot-rendering contract that uses existing visualization logic without thermodynamic backend calls;
• guarded no-overwrite promotion of one image and provenance sidecar into a worker-derived directory under the workspace figures/ root;
• Plot-page format selection, manual Render controls, row/advisory reporting, and an informational equivalent CLI command;
• immediate confirmed force-stop with parent-owned staging cleanup and guarded close behavior;
• validated Qt-only PNG/SVG previews with fit, zoom, 100%, and panning controls; and
• explicit PDF opening through the system viewer.

Imported invalid YAML files remain untouched and must be repaired in a text editor before import. Imported valid files also remain untouched until saved under the selected workspace. Exact numeric visualization filters and series levels remain explicit inputs; finite choices such as fields, fluids, categorical values, units, formats, scales, and valid series dimensions are provided by the editor.

Sweep and preparation creation remain reserved for GUI-2; GUI-1 inspects their completed bundles read-only. NumPy and SafeTensors outputs are listed from the preparation manifest but are not rendered as matrices. The Plot page can build a compatible request from inspection results and render PNG, SVG, or PDF through the worker. PNG and SVG exports preview automatically after containment, symlink, regular-file, suffix, and SHA-256 checks. PDF exports remain closed until the user selects Open PDF and the file passes the same checks again. Preview failures do not remove successful image or sidecar exports.

Qt/PySide6 is an optional third-party dependency with its own LGPL/GPL/commercial licensing terms. Carnopy remains MIT licensed, does not vendor Qt, and does not distribute standalone desktop installers in GUI-1. Downstream redistributors should review the official Qt for Python package details and Qt licensing terms. This is not legal advice.

Architecture map

The repository includes a generated Graphify codebase map:

graphify-out/GRAPH_REPORT.md
graphify-out/graph.html
graphify-out/graph.json

Open graphify-out/graph.html locally after cloning for an interactive view. The graph is an aid for navigation and review, not a source of scientific truth. Verify exact behavior against the source files and tests before making changes.

Scientific limitations

• CoolProp is the only backend in Milestone 1.
• CoolProp model selection supports HEOS, Peng-Robinson, and Soave-Redlich-Kwong.
• Pure fluids only; mixtures are deferred.
• Generated data is backend output, not experimental evidence.
• All backend calls and generated numeric columns use SI.
• Specific enthalpy, entropy, and internal energy depend on reference state.
• Carnopy resets every requested fluid to CoolProp DEF before generation and records that policy.
• Preparation warns and records reference-state context when selected features or targets include absolute reference-dependent values.
• CoolProp reference-state mutation is process-global; concurrent embedded use with unrelated CoolProp calculations is unsupported in Milestone 1.
• Release regression tests compare finalized Parquet values with direct CoolProp calls for representative states in all three modes.
• Separate sanity checks require the generated normal boiling points of Propane and Cyclopentane at 101325 Pa to remain within the uncertainty intervals published by the NIST Chemistry WebBook. These checks do not establish universal experimental accuracy.
• Absolute reference-dependent values are not directly comparable across different reference conventions or model/reference combinations.
• PR/SRK transport properties, surface tension, and triple-point temperature are rejected during validation because CoolProp 7.2.0 does not provide the required model capability.
• Visualization reads emitted columns only and is not a second property evaluation layer.
• ORC generation, additional backends, ML training, web services, databases, and mixture models are deferred.

Post-alpha work may add an optional cycle-feasibility subsystem that produces traceable screening datasets without turning the property generator into a hidden process simulator. An ORC/TFC contract must explicitly include source and sink profiles, pinch/approach temperatures, pressure losses, component efficiencies, subcooling and superheat margins, cavitation/NPSH constraints, minimum turbine-exhaust quality, and critical/maximum operating limits. Saturated liquid alone is not a pump cavitation margin, and turbine discharge need not universally have vapor mass fraction one.

Official backend references:

• https://coolprop.org/coolprop/
• https://coolprop.org/coolprop/HighLevelAPI.html
• https://github.com/CoolProp/CoolProp

Development and contribution

Carnopy uses a src/ layout, Hatchling, standalone uv, Ruff, strict mypy, and pytest. pyproject.toml and uv.lock are authoritative.

Normal development:

uv sync --locked --extra all --group dev

Release-readiness tooling:

uv sync --locked --extra all --group dev --group release

Quality gate:

uv lock --check
uv run --locked ruff check .
uv run --locked ruff format --check .
uv run --locked mypy src/carnopy
uv run --locked pytest
uv run --locked python scripts/preflight.py
uv pip check --python .venv/bin/python

Full local release gate, including a non-destructive build and distribution inspection:

bash scripts/local_gate.sh prerelease/local-gate

Keep changes small and explicit. Public configuration names, semantic property names, SI dataset columns, failure codes, metadata fields, and identity rules are compatibility contracts. Tests use temporary output directories and do not commit generated datasets or figures.

The test count is not a quality target. The suite separates configuration, sampling, three thermodynamic modes, diagnostics, provenance, visualization, CLI behavior, packaging, and release automation. New tests should protect a distinct contract or regression and use parametrization instead of duplicating equivalent cases.

Contributor and coding-agent rules, architecture constraints, commit conventions, and release-maintainer safeguards are in AGENTS.md. Contributor setup, testing, and pull-request guidance are in CONTRIBUTING.md. Report security vulnerabilities privately according to the security policy.

Project status and roadmap

Carnopy remains alpha software while its public schemas and backend boundaries are validated through real use. Version 0.1.0a3 includes GUI-1, a Linux-first desktop frontend for the existing dataset workflow.

GUI-1 includes the worker protocol, workspace shell, dataset configuration editor, dataset execution, read-only bundle inspection, bounded table previews, job diagnostics, guarded staging recovery, inspection-driven plot requests, worker-side backend-free rendering, safe no-overwrite artifact promotion, desktop Render controls, confirmed force-stop/close ownership, Qt-only PNG/SVG previews, explicit PDF opening, packaging inventory checks, installed-wheel smoke coverage, release hardening, and final architecture-map review.

The prepared source now includes robust distribution reports, exact-state split leakage protection, explicit-bin stratification, train-fitted robust scaling, opt-in matrix diagnostics, and optional diagnostic-only scikit-learn baselines. The ML preparation roadmap separates current behavior from research directions. Carnopy remains a dataset and preparation tool rather than a model-training framework.

A later pure-fluid ORC feasibility-envelope subsystem may produce traceable accepted and rejected operating windows rather than silently acting as a complete process simulator or optimizer.

That design must explicitly cover source and sink profiles, pinch and approach temperatures, pressure losses, subcooling and superheat margins, equipment efficiencies, critical-point and operating limits, and minimum turbine-exhaust quality. Saturated liquid alone is not a pump cavitation margin; NPSH may be reported only when sufficient hydraulic-system and pump data are supplied.

The planned 0.1.0a4 application milestone is a cross-platform modern QML presentation layer with optional native VTK exact-grid 3D visualization. GUI-2 will use the same worker and core Python boundaries rather than duplicate scientific logic. TFC screening, mixtures, additional backends, and standalone desktop installers remain deferred.

Use GitHub issues for bug reports, scientific discrepancies, and focused feature requests. See CONTRIBUTING.md before proposing a public or scientific contract change.

License

Carnopy is distributed under the MIT License. See LICENSE.