acropole 0.1.2

Open-source library to predict aircraft fuel flow from trajectory data using machine learning models

Acropole — Predict aircraft fuel flow from trajectory data

---

acropole predicts the fuel flow of aircraft (kg/s, kg/h and cumulative kg) from trajectory data — groundspeed, altitude and vertical rate — using a portable ONNX model trained on Quick Access Recorder (QAR) data. It accepts a pandas or polars DataFrame, dispatches per aircraft typecode, and returns the same frame enriched with fuel-flow columns. The runtime depends only on numpy + polars + onnxruntime — no heavy ML framework.

📖 Full documentation

Features

• ⛽ Fuel-flow prediction — per-sample fuel_flow (kg/s), fuel_flow_kgh (kg/h) and fuel_cumsum (kg) from a flight trajectory
• ✈️ Multi-aircraft — frames mixing several typecodes are scored per typecode, each row with its own aircraft parameters
• 🐼 pandas and polars — estimate() accepts either and returns the same type; the engine runs on polars internally
• 🚀 Fast ONNX runtime — migrated from TensorFlow, 2–4.8× faster depending on batch size, numerical parity validated to 1e-6; no TensorFlow dependency
• 📈 Temporal derivatives — supply a second column to compute accelerations (and fuel_cumsum), or pass pre-computed derivatives directly
• 🎯 Column mapping — map your own column names with keyword arguments, no renaming required
• ⚡ Typecode-bound fast path — AircraftFuelEstimator offers numpy-only I/O with pre-computed parameters for hot loops
• 🧪 Fully typed — ships py.typed, strict mypy on source and tests, 98% coverage

Installation

acropole is published on PyPI:

uv add acropole      # or: pip install acropole

The core library runs on numpy + polars + onnxruntime. To pass and receive pandas DataFrames, install the optional pandas extra:

uv add "acropole[pandas]"      # or: pip install "acropole[pandas]"

To work from source, this project uses uv:

git clone https://github.com/DGAC/Acropole.git
cd Acropole
uv sync

Quick Start

import pandas as pd
from acropole import FuelEstimator

fe = FuelEstimator()

flight = pd.DataFrame({
    "typecode": ["A320", "A320", "A320", "A320"],
    "groundspeed": [400, 410, 420, 430],     # kt
    "altitude": [10000, 11000, 12000, 13000],  # ft
    "vertical_rate": [2000, 1500, 1000, 500],   # ft/min
    # optional features:
    "second": [0.0, 4.0, 8.0, 12.0],
    "airspeed": [400, 410, 420, 430],
    "mass": [60000, 60000, 60000, 60000],
})

flight_fuel = fe.estimate(flight)
# adds fuel_flow (kg/s), fuel_flow_kgh (kg/h), fuel_cumsum (kg)

Notes:

• Providing the second column makes the estimation more accurate (it derives accelerations of the speeds) and enables fuel_cumsum.
• airspeed is optional; if absent it defaults to groundspeed. Accurate airspeed improves the estimate.
• The expected sampling rate is 4 seconds. A higher or lower rate produces noisier fuel flow — resample before estimating.

For a full notebook, see examples/fuel_estimation.ipynb.

Advanced Usage

Mapping your own column names

estimate() accepts keyword arguments mapping each logical feature to the column name in your frame. The example flight (QAR columns) maps like this:

import pandas as pd
from acropole import FuelEstimator

flight = pd.read_csv("examples/example_flight.csv")
flight = flight.iloc[::4]  # resample to ~4 s

fe = FuelEstimator()
flight_fuel = fe.estimate(
    flight,
    typecode="FLPL_AIRC_TYPE",
    groundspeed="GRND_SPD_KT",
    altitude="ALTI_STD_FT",
    vertical_rate="VERT_SPD_FTMN",
    second="FLIGHT_TIME",
    airspeed="TRUE_AIR_SPD_KT",
    mass="MASS_KG",
)

Required logical columns: typecode, groundspeed (kt), altitude (ft), vertical_rate (ft/min). Optional: second, airspeed, mass, and the pre-computed derivatives d_altitude / d_groundspeed / d_airspeed.

Multi-aircraft frames

A single frame can mix typecodes; each row is scored with its own aircraft parameters. Unsupported typecodes emit a warning and leave fuel_flow as NaN:

flight = pd.DataFrame({
    "typecode": ["A320", "A320", "B738", "B738"],
    "groundspeed": [400, 410, 420, 430],
    "altitude": [10000, 11000, 12000, 13000],
    "vertical_rate": [2000, 1500, 1000, 500],
})
fe.estimate(flight)  # A320 rows and B738 rows each use their own params

Typecode-bound fast path

For hot loops on a single aircraft, AircraftFuelEstimator works directly on numpy arrays (no per-call params lookup) and returns fuel flow in kg/s:

import numpy as np
from acropole import FuelEstimator

# share the already-loaded ONNX session and parameters
afe = FuelEstimator().for_aircraft("A320")

fuel_flow = afe.estimate(
    groundspeed=np.array([400.0, 410.0, 420.0]),
    altitude=np.array([10000.0, 11000.0, 12000.0]),
    vertical_rate=np.array([2000.0, 1500.0, 1000.0]),
)  # kg/s  (multiply by 3600 for kg/h)

Custom aircraft parameters and model

fe = FuelEstimator(
    aircraft_params_path="path/to/your/aircraft_params.csv",
    model_path="path/to/your/model.onnx",
)

Aircraft parameters from open data are bundled in src/acropole/data/aircraft_params.csv and the ONNX model in src/acropole/models/; both are loaded by default.

Command-Line

Installing acropole also installs the acropole command (entry point acropole.cli:main). It reads a flight from a CSV or parquet file, estimates fuel flow, and writes the enriched table back to disk — adding fuel_flow (kg/s), fuel_flow_kgh (kg/h) and, with --second, fuel_cumsum (kg):

acropole estimate examples/example_flight.csv \
  --typecode FLPL_AIRC_TYPE --groundspeed GRND_SPD_KT --altitude ALTI_STD_FT \
  --vertical-rate VERT_SPD_FTMN --airspeed TRUE_AIR_SPD_KT --mass MASS_KG \
  --second FLIGHT_TIME --out result.csv
# wrote 7796 rows with fuel columns to result.csv

If your file already uses the standard column names, no mapping is needed and the output defaults to <flight>_fuel.<ext> next to the input:

acropole estimate flight.csv        # writes flight_fuel.csv

Each --typecode, --groundspeed, --altitude, --vertical-rate, --airspeed, --mass and --second flag maps a logical feature to the matching column in your file (defaults are the standard names). See the CLI reference for all options.

Comparison of Different Model Performances

Comparison of different model performances per phase for 1000 test flights of A320-214 aircraft using real mass and true airspeed.

Phase	Samples #	Metric	ACROPOLE	OpenAP	OpenAP V2	BADA	Poll-Schumann
		MAPE (%)	2.13	30.35	8.84	6.53	6.85
CLIMB	1,403,850	MAE (kg/min)	1.66	25.81	6.92	5.53	5.65
		ME (kg/min)	0.85	-25.66	-2.48	-5.27	-4.62
		MAPE (%)	4.41	18.59	10.69	7.01	4.84
LEVEL	4,017,801	MAE (kg/min)	1.82	7.82	3.48	2.65	2.03
		ME (kg/min)	1.22	-7.47	2.64	-1.43	-0.73
		MAPE (%)	12.63	51.69	32.4	21.50	21.55
DESCENT	1,684,117	MAE (kg/min)	2.71	8.62	5.58	3.71	4.71
		ME (kg/min)	1.88	-1.75	-1.16	-0.64	-3.67
		MAPE (%)	5.91	27.60	14.71	9.84	8.61
ALL	7,105,768	MAE (kg/min)	1.99	11.55	4.58	3.44	3.29
		ME (kg/min)	1.30	-9.92	0.84	-2.03	-2.09
		Processing time (s)	3	284	255	474	28

The Acropole model is a neural network trained on Quick Access Recorder (QAR) data from several aircraft types. Evaluation and the list of supported aircraft are available at evaluation/Dense_Acropole_FuelFlow_Scaling.

Development

Set up a full development environment (runtime, dev, and docs groups):

uv sync --all-groups

Common tasks:

uv run pytest                    # run the test suite (98% coverage)
make lint                        # ruff + mypy strict on src AND tests
pre-commit install               # enable the pre-commit hooks

License

Acropole is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0). See the LICENSE file for the full text.

Credits

@inproceedings{jarry2024towards,
  title={Towards aircraft generic Quick Access Recorder fuel flow regression models for ADS-B data},
  author={Jarry, Gabriel and Delahaye, Daniel and Hurter, Christophe},
  booktitle={International Conference on Research in Air Transportation},
  year={2024}
}