mfbo 0.1.1

Multi-fidelity surrogate models (NN ensembles and co-kriging) compatible with BoTorch.

mfbo

"mfbo" is a Python library for "multi-fidelity surrogate modeling and Bayesian optimization". It provides neural-network ensemble surrogates and Gaussian-process co-kriging models that are fully compatible with "BoTorch". The library is designed for scientific and engineering optimization problems.

mfbo is a Python library for multi-fidelity surrogate modeling and Bayesian optimization. It provides neural-network ensemble surrogates and Gaussian-process co-kriging models that are fully compatible with BoTorch.

The library is designed for scientific and engineering optimization problems where high-fidelity evaluations are expensive and multiple fidelity levels are available.

---

Reference

This library implements methods described in:

Passakorn Paladaechanan et al.
Adaptive Gated Multi-Fidelity Neural Networks for Bayesian Optimization
Ocean Engineering, 2025
https://www.sciencedirect.com/science/article/pii/S002980182502997X

If you use this library in academic work, please cite the above paper.

---

Features

• Multi-fidelity neural surrogate ensembles
  • MLPEnsemble
  • MFNNEnsemble
  • AGMFNetEnsemble
  • Ada2MFEnsemble
• Gaussian-process surrogate models
  • Kriging
  • AR(1) Co-Kriging (Kennedy–O’Hagan model)
• Native BoTorch compatibility
  • All models expose a posterior() method
• Supports multi-output objectives
• Designed for Bayesian optimization workflows

---

Dependencies

mfbo is built on the PyTorch Bayesian optimization ecosystem:

• PyTorch ≥ 2.5
• GPyTorch ≥ 1.15
• BoTorch ≥ 0.16
• tqdm ≥ 4.67

Please install PyTorch separately according to your system (CPU or CUDA) before installing mfbo.

Installation

1. Install PyTorch

Install PyTorch first (CPU or CUDA version depending on your system):

https://pytorch.org/get-started/locally/

2. Install mfbo

pip (PyPI)

pip install mfbo

conda (conda-forge)

conda install -c conda-forge mfbo

Contact

For questions related to the implementation or the associated research,
please contact:

Passakorn Paladaechanan
Email: [e-mail]