MFTT enables the construction of monotone triangular transport maps from multifidelity training data. It implements the single-fidelity, hierarchical, and nonhierarchical constructions in the manuscript Multifidelity Formulations for Triangular Transport (arXiv:2610.00698, 2026), with particular attention to settings where high-fidelity data are scarce.
MFTT supports Python 3.10 or newer. Install it from a repository clone:
git clone https://github.com/sandialabs/multifidelity-triangular-transport.git
cd multifidelity-triangular-transport
python3 -m venv .venv
source .venv/bin/activate
python -m pip install .The installation guide includes verification instructions.
import numpy as np
from mftt import MapParams, TriangularMap
rng = np.random.default_rng(2026)
u, v = rng.standard_normal((2, 600))
samples = np.column_stack((u, v + 0.5 * (u**2 - 1.0)))
model = TriangularMap(samples, MapParams(total_order=2))
model.train()
reference_samples = model.evaluate(samples)
recovered_samples = model.inverse(reference_samples[:32])
np.testing.assert_allclose(recovered_samples, samples[:32], atol=1e-8)See the single-fidelity guide for the full model workflow and interpretation.
The complete rendered documentation is available at sandialabs.github.io/multifidelity-triangular-transport.
The manuscript is the authority for the mathematical definitions and method names. See the citation page for software authorship and the manuscript reference.
Contributor setup and local documentation build commands are in the developer guide.
MFTT is developed by Owen Davis and Gianluca Geraci and released under the MIT License. The implemented methods are based on the referenced manuscript by Owen Davis, Daniel Sharp, Youssef Marzouk, and Gianluca Geraci.