opentop: Open Trajectory Optimizer

opentop is v2 of openap-top. The package has been renamed and restructured: what used to be installed as openap-top and imported as openap.top now installs as opentop and imports as import opentop. The three headline changes in v2.0:

• New Opti-stack backend — NLP construction has moved to CasADi's Opti stack, removing ~400 lines of boilerplate and fixing several long-standing bugs.
• Standalone package — opentop is no longer a namespace extension of openap; it installs as a top-level package and imports as import opentop.
• Command-line interface — opentop optimize and opentop gengrid expose the optimizer and grid-cache builder from the shell.

See What's New in 2.0 below for the full migration table.

Flight trajectory optimizer based on the OpenAP aircraft performance model.

opentop uses non-linear optimal control via direct collocation (CasADi + IPOPT) to generate optimal flight trajectories. It provides simple interfaces for:

• Complete flight trajectories (takeoff → cruise → landing)
• Individual phases: climb, cruise, descent
• Fuel-optimal, time-optimal, cost-index, and climate-optimal objectives
• Wind integration
• Custom 4D grid cost functions (contrails, weather, airspace)
• User-defined objective functions and constraints

🕮 User Guide

Detailed guide and examples: https://openap.dev/optimize.

Install

From PyPI:

pip install --upgrade opentop

From the development branch:

pip install --upgrade git+https://github.com/junzis/openap
pip install --upgrade git+https://github.com/junzis/opentop

opentop is a standalone package. Prior to v2.0 it shipped as openap.top, a namespace extension of openap; v2.0 drops that and installs as a top-level opentop package instead.

Control representation

OpenTOP 2.6 uses continuous, piecewise-linear Mach, vertical-speed and heading controls. With nodes=N intervals it optimizes N+1 boundary controls, including an independent arrival control. Dynamics and objective quadrature evaluate the interpolated control at every collocation point. Phase rate limits cover every interval using its actual duration, including variable timesteps. Cruise performance constraints also cover internal collocation points; nonlinear feasibility still requires between-node checks. For CompleteFlight, trajectory(path_constraint_points=[(interval, tau), ...]) adds force and energy constraints at selected points inside intervals. Interval indices start at zero; tau is the local time fraction from 0 to 1. States use the collocation polynomial and controls use linear interpolation. Select these points from a path audit for the specific route, mesh, and objective, then audit the new solution again. The optimizer does not select them automatically, and these additional samples do not guarantee feasibility everywhere between them.

This replaces the previous piecewise-constant control transcription. Saved trajectory DataFrames retain their columns and N+1 rows. The numeric export helper still accepts historical N-column control arrays for replaying old data. Continuity does not imply continuous control derivatives, a certified continuous path envelope, or a complete acceleration-energy model.

Quick Start

A simple optimal flight

import opentop

optimizer = opentop.CompleteFlight("A320", "EHAM", "LGAV", m0=0.85)
flight = optimizer.trajectory(objective="fuel")

flight is a Pandas DataFrame with columns for position, altitude, mass, Mach, TAS, vertical rate, heading, and per-segment fuel_cost.

Other built-in objectives

optimizer.trajectory(objective="time")      # minimum time
optimizer.trajectory(objective="ci:30")     # cost index 30
optimizer.trajectory(objective="gwp100")    # 100-yr global warming potential
optimizer.trajectory(objective="gtp100")    # 100-yr global temperature potential

The supported climate metrics are gwp20, gwp50, gwp100, gtp20, gtp50, gtp100.

Choosing a different engine

optimizer = opentop.CompleteFlight(
    "A320", "EHAM", "LGAV", m0=0.85, engine="CFM56-5B4"
)

Flight phase optimizers

cruise    = opentop.Cruise("A320", "EHAM", "LGAV", m0=0.85).trajectory()
climb     = opentop.Climb("A320", "EHAM", "LGAV", m0=0.85).trajectory()
descent   = opentop.Descent("A320", "EHAM", "LGAV", m0=0.85).trajectory()

Cruise also supports constant-altitude, constant-Mach, and fixed-track modes:

opt = opentop.Cruise("A320", "EHAM", "LGAV", m0=0.85)
opt.fix_cruise_altitude()
opt.fix_mach_number()
opt.fix_track_angle()
flight = opt.trajectory()

Multiple-aircraft separation optimization

MultiAircraft formulates several cruise trajectories in one CasADi Opti problem and applies the default 5 NM / 1,000 ft high-order separation model:

import opentop

flights = [
    opentop.FlightSpec(
        "AC1", opentop.Cruise("A320", (52.3, 9.8), (49.5, 2.5)), start_time=0
    ),
    opentop.FlightSpec(
        "AC2", opentop.Cruise("B738", (52.9, 4.6), (49.5, 9.3)), start_time=140
    ),
]

result = opentop.MultiAircraft(flights).trajectory()
df_ac1 = result.trajectories["AC1"]

The default constraint is (dx / Rxy)^2 + (dy / Rxy)^2 + (dh / Rz)^8 >= 1.3. Pass a SeparationConfig only to customize it. Fleet DataFrames preserve relative ts and add absolute_ts using each flight's configured start time.

To study horizontal conflict avoidance at one shared flight level, first call fix_cruise_altitude() on every child optimizer and then pass common_altitude=True. The common altitude remains an optimization variable:

for flight in flights:
    flight.optimizer.fix_cruise_altitude()

result = opentop.MultiAircraft(flights, common_altitude=True).trajectory()

Descent optimizers can also be coordinated. List flights in the desired landing order and use minimum_arrival_gap_s to enforce a runway sequence. Shared final-approach points can be supplied with ordered waypoint constraints. For structured arrivals, route_heading_tolerance_deg keeps each route leg near its nominal bearing, while max_duration_s can impose an operational descent-time bound without adding time to the objective. See the coordinated-arrivals notebook for a three-aircraft runway-arrival example.

The implementation supports uniform-timestep Cruise and Descent optimization, including ordered waypoint constraints. Wind, variable timesteps, time-dependent 4D grid objectives, and other phases are rejected explicitly until their fleet timing semantics are implemented.

Wind integration

Download ERA5 (or similar) meteorological data in GRIB format, then:

import opentop

windfield = opentop.tools.read_grids("wind.grib")

optimizer = opentop.CompleteFlight("A320", "EHAM", "LGAV", m0=0.85)
optimizer.enable_wind(windfield)

flight = optimizer.trajectory(objective="fuel")
opentop.vis.trajectory(flight, windfield=windfield, barb_steps=15)

Custom grid cost (contrails, weather, airspace)

Build a CasADi interpolant from a DataFrame with columns longitude, latitude, height (m), cost, and optionally ts:

interpolant = opentop.tools.interpolant_from_dataframe(df_cost)

def contrail_objective(x, u, dt, **kwargs):
    grid_cost = optimizer.obj_grid_cost(
        x, u, dt, interpolant=kwargs["interpolant"], n_dim=3
    )
    fuel_cost = optimizer.obj_fuel(x, u, dt)
    return grid_cost + fuel_cost

flight = optimizer.trajectory(
    objective=contrail_objective,
    interpolant=interpolant,
    n_dim=3,
)

See https://openap.dev/optimize/contrails.html for a full contrail + CO₂ example.

Custom objective functions

Any callable with signature (x, u, dt, **kwargs) -> ca.MX can be used:

def my_objective(x, u, dt, **kwargs):
    # x: state [xp, yp, h, mass, ts]
    # u: control [mach, vs, psi]
    return your_cost_expression

flight = optimizer.trajectory(objective=my_objective)

Multi-start optimization for non-convex objectives

Trajectory optimization with grid costs, blended objectives, or tight constraints can have multiple local minima — a single solve lands in whichever basin is closest to the initial guess. For problems where robustness matters, multi_start_trajectory runs N solves from different randomized initial guesses and returns the best:

trajectory, candidates = optimizer.multi_start_trajectory(
    objective=contrail_objective,
    interpolant=interp,
    max_fuel=6500,
    n_starts=5,
    lateral_jitter_km=100.0,
    altitude_jitter_ft=3000.0,
    seed=0,
)

Start 0 uses the canonical initial guess (your initial_guess= if provided, otherwise the default great-circle). Starts 1..N-1 are random perturbations of the canonical: sinusoidal lateral bulges (endpoints preserved) and constant altitude offsets.

Returned is a (trajectory, candidates) tuple. trajectory is the winning DataFrame (feasibility-first, then lowest objective). candidates is a best-first ordered list of dicts with per-start metadata (objective, fuel, grid_cost, success, iters, perturbation, wall_time_s, trajectory).

n_starts=1 is identical to calling trajectory() directly, so adding multi_start_trajectory to existing code is additive.

Precomputed grid caches (recommended for contrail + CO₂)

Linear interpolation over a 4D contrail-cost grid has discontinuous derivatives at every grid cell boundary, which can cause IPOPT's line search to oscillate on non-convex blended objectives. The fix is to use a cubic B-spline interpolant, which has continuous derivatives. But building a bspline over a large grid can take several minutes, so we expose a cache utility:

from opentop.tools import cached_interpolant_from_dataframe

interpolant = cached_interpolant_from_dataframe(
    df_cost, "cache/contrail.casadi", shape="bspline"
)

First call builds the bspline and writes it to disk (~1-3 minutes for a 60k-point slice); subsequent calls load the cache in under a second.

If your grid only covers the altitude band where contrails actually form (typically FL200-FL440), extend it with zero-cost levels outside that band before building the interpolant — otherwise opentop.CompleteFlight trajectories that start and end on the ground will query the interpolant outside its data range. The opentop CLI has a helper for this:

opentop gengrid --in raw_grid.parquet --out grid.casadi \
    --bbox 35:57,-9:7 --time 2022-02-20T10:00,2022-02-20T14:00 \
    --pad-altitudes

--pad-altitudes is on by default and adds zero-cost rows at altitudes from 0 to FL480 so the interpolant returns 0 (physically correct — no contrails below ~FL200 or above ~FL440) outside the data band.

Command-line interface

Installing opentop also installs the opentop executable, which exposes two subcommands: optimize (run a trajectory optimization) and gengrid (precompute a grid-cost interpolant).

opentop optimize

Run a trajectory optimization without writing any Python:

opentop optimize EHAM EDDF -a A320 --phase cruise --obj fuel

A concise solver summary (status, iterations, objective, fuel burn, max altitude, flight time) is printed to stdout. Pass -o flight.parquet to also save the full trajectory DataFrame.

Supported objectives: fuel, time, ci:N (cost index, any integer), gwp20/gwp50/gwp100, gtp20/gtp50/gtp100, and grid (requires --grid FILE).

Blended objectives are written as a weighted sum:

opentop optimize EHAM EDDF -a A320 --phase all \
    --obj "0.3*fuel+0.7*grid" \
    --grid contrail.casadi

Common flags:

opentop gengrid

Build and cache a CasADi interpolant from a raw cost grid:

opentop gengrid --in raw_grid.parquet --out contrail.casadi \
    --bbox 35:57,-9:7 \
    --time 2022-02-20T10:00,2022-02-20T14:00 \
    --shape bspline

The resulting .casadi file loads in under a second (vs. minutes to rebuild a bspline from raw grid data), so keep it on disk and pass it to opentop optimize --grid.

Use opentop --help, opentop optimize --help, and opentop gengrid --help for the full option list.

opentop replay

Replay a real flight by callsign: pulls trajectory from OpenSky, ERA5 meteo from fastmeteo, and runs an optimization side-by-side with the actual flight.

# Install with the replay extras
pip install "opentop[replay]"

# Flag mode
opentop replay RYR880W --date 2023-01-05 --obj fuel -o ./results/

# Interactive wizard (no arguments)
opentop replay

Outputs under the output directory:

• actual.parquet — cleaned OpenSky trace
• optimized.parquet — optimized trajectory
• trajectory.png — actual-vs-optimized overlay

Requires OpenSky credentials at ~/.config/traffic/traffic.conf (see traffic's docs). ERA5 data is fetched via fastmeteo's ArcoEra5; the default local store is <system tmp>/opentop-era5 — first use will download several GB on a large bbox.

Accessing Solver Results

After calling .trajectory(), the optimizer exposes:

optimizer.stats           # solver statistics dict ("success", "iter_count", ...)
optimizer.success         # True if the most recent solve succeeded
optimizer.objective_value # final objective value (float)

For full structured access including status, iteration count, fuel, and grid cost in one dataclass:

result = optimizer.trajectory(objective="fuel", result_object=True)
# result.df, result.success, result.status, result.objective,
# result.iters, result.fuel, result.grid_cost, result.stats

optimizer.solver still works in v2.2 with a DeprecationWarning; it will be removed in v2.3.

Benchmarks

Run benchmarks across versions to verify performance:

./benchmark.sh                 # Benchmark HEAD (local dev code)
./benchmark.sh v2.0.0          # Benchmark a specific PyPI release
./benchmark.sh v1.11.0 v2.0.0  # Benchmark multiple versions sequentially

Reports land in tests/benchmarks/<version>.txt.

Migrating to 2.2

Version 2.2 makes several structural changes that may require small updates to existing code:

Internally, opentop/ has been split into focused modules: _dynamics.py, _objectives.py, _trajectory.py, _options.py, _multi_start.py. If you were importing internal helpers, they have moved — use the public API on Base / Cruise / CompleteFlight where possible.

MultiPhase was rarely used and its functionality is fully covered by CompleteFlight. If you had code using MultiPhase, replace it with:

full = opentop.CompleteFlight("A320", "EHAM", "LGAV", m0=0.85).trajectory()

optimizer.solver still works in v2.2 with a DeprecationWarning. It will be removed in v2.3 — prefer optimizer.stats / optimizer.success.

Type annotations are now enforced in CI via pyright (basic mode). Public API signatures are fully annotated; see opentop/_options.py for the new SolveOptions, GridOptions, and TrajectoryResult dataclasses.

What's New in 2.0

Version 2.0 is a major refactor. Most user code keeps the same shape, but a few things have moved:

The NLP construction moved to CasADi's Opti stack, which removed ~400 lines of boilerplate and cleaned up several bugs. The module rename from openap.top to opentop eliminates the namespace-extension install mode that used to require .pth tricks.

See the changelog for details.

License

MIT License. See LICENSE for details.