AeroOpt (aeroopt)

A multi-objective (and single-objective) optimization framework for engineering workflows where evaluating a design is the expensive part — a CFD run, an FEA solve, or any external executable that takes minutes to hours per sample.

That premise drives the design:

• Every evaluation is kept. Designs accumulate in a persistent Database that can be written to JSON/Excel, reloaded, merged, sub-setted and restarted from. The optimizer works on an archive, not a transient generation.
• Evaluation is pluggable and parallel. Objectives can be Python callables or external solvers driven in their own working folders; MultiProcessEvaluation spreads a generation across processes.
• The loop is open. PreProcess / PostProcess hooks let you repair, screen or replace candidates before an expensive evaluation is spent on them.
• Failure is expected. Diverged solves and constraint violations are recorded rather than dropped, and the search adapts to how much feasible data actually exists.

Features

• Problems and data: Problem, Individual, Database; constraint strings and custom constraint callables; JSON / Excel serialization.
• Evaluation: built-in Python objectives or external executables; MultiProcessEvaluation for parallel runs (Linux and Windows).
• Optimization loop: OptBaseFramework with pluggable PreProcess / PostProcess.
• Evolutionary algorithms: NSGA-II, NSGA-III, RVEA, MOEA/D, differential evolution (MODE-style) and NRBO.
• Surrogates and hybrids: SAO and SBO in aeroopt.optimization.hybrid; aeroopt.utils.surrogate defines the surrogate interface (Kriging via SMT).
• Analysis: AnalyzeDatabase for input-space crowding metrics, potential fields and clustering; standard test functions in aeroopt.utils.benchmark.

For richer visualization and decision-making around a computed Pareto set, see pymoo.

Requirements

• Python ≥ 3.9
• numpy, scipy, scikit-learn, numexpr, pydoe>=0.9.8, openpyxl

Installation

pip install aeroopt                 # core
pip install "aeroopt[surrogate]"    # + smt, for Kriging
pip install "aeroopt[examples]"     # + matplotlib, for the example scripts

Editable install from a clone:

pip install -e ".[surrogate,examples,tests]"

Quick start

import numpy as np
from aeroopt.core import Problem, SettingsData, SettingsProblem
from aeroopt.optimization import OptNSGAII, SettingsOptimization, SettingsNSGAII

def evaluate(x: np.ndarray):
    """Return (succeed, y); succeed=False records a failed evaluation."""
    return True, np.array([x[0], 1.0 - np.sqrt(x[0]) + x[1]])

data_settings = SettingsData.from_values(
    'demo_data',
    name_input=['x1', 'x2'], input_low=[0.0, 0.0], input_upp=[1.0, 1.0],
    name_output=['f1', 'f2'], output_low=[-0.1, -1.0], output_upp=[1.1, 10.0],
)

problem_settings = SettingsProblem.from_values(
    'demo_problem', data_settings,
    output_type=[-1, -1],                        # both minimized
    constraint_strings=['x1 ** 2 + x2 ** 2 - 0.64'],
)

opt = OptNSGAII(
    problem=Problem(data_settings, problem_settings),
    optimization_settings=SettingsOptimization.from_values(
        'demo_opt', population_size=32, max_iterations=20, seed=42),
    algorithm_settings=SettingsNSGAII.from_values('demo_alg'),
    user_func=evaluate,
)
opt.main()

print(opt.db_elite.size, 'non-dominated designs')

Configuration: Python or JSON

The settings above are defined inline. The same objects can equivalently be read from a JSON file, which is the better choice for a study worth version-controlling:

data_settings = SettingsData('demo_data', fname_settings='settings.json')

save_settings([...], 'settings.json') converts a Python-defined study into that file, and it reads back unchanged. See aeroopt/template_settings.json for a template with every supported entry.

Documentation

pip install -e ".[docs]"
sphinx-build -b html docs/source docs/build/html

The documentation covers the settings reference, the architecture, and the principles behind each algorithm (dominance and crowding, SBX/polynomial mutation, DE, NRBO, NSGA-III niching, RVEA's angle-penalized distance, MOEA/D decomposition, and surrogate-driven optimization).

Package layout

Algorithms

Examples (example/)

Scripts prepend the repository root to sys.path so they run from a clone without installing; remove that block if you installed the package.

Tests

pytest

Contributing

Development principles, architecture notes and a change checklist are in AGENTS.md — written for humans and AI agents alike.

An agent skill for using the library lives in skills/aeroopt/: point your assistant at it, or read skills/aeroopt/SKILL.md yourself for a condensed guide to the settings, the algorithms and the common pitfalls.

Repository

https://github.com/swayli94/AeroOpt