mille-feuille

mille‑feuille acts as an orchestrator when running sampling, learning and optimisation loops against expensive MPI-parallelised HPC codes. For optimisation, mille‑feuille is a thin wrapper on top of BoTorch, providing the necessary interface between simulators, surrogates and optimisers.

Status: early days – very much a work in progress

Used in the following publications:

• Automated simulation-based design via multi-fidelity active learning and optimisation for laser direct drive implosions, Crilly et al.

🔧 Install

Pip:

```bash

Development head

pip install git+https://github.com/aidancrilly/mille-feuille.git ```

Or clone repo and install dev environment:

bash git clone https://github.com/aidancrilly/mille-feuille.git cd mille-feuille pip install -e .[dev] python dev_fetch_deps.py # Grabs header file needed by C++ test script

Requires Python ≥ 3.11. Core dependencies (botorch, gpytorch, numpy, scipy, h5py, scikit‑learn …) are pulled in automatically.

🚀 Quick‑start

Take a look at the examples directory and sub-directories within:

• test_executables: simple (fortran90 and C++) examples implemented for the test suite.
• loops: example scripts which use mille‑feuille to perform sampling, learning and optimisation tasks. This example includes a template for a simulator with namelist based input and scheduling within a PBS environment.

Core components

Domains, States and Surrogates

mille‑feuille implements the following containers:

1. InputDomain holds the bounded input domain which can be a mix of continuous and discrete dimension
2. FidelityDomain holds information regarding the degrees of simulation fidelity.
3. State holds the necessary data taken from simulation samples: indices (Is), inputs (Xs), output parameters (Ps), fidelities (Ss) and objectives (Ys)

These classes hold the necessary information to train surrogate models. mille‑feuille has a number of abstract base classes as well as concrete examples of surrogate models including Gaussian Processes (using GPyTorch and BOTorch) and Neural Network Ensembles (using PyTorch and BOTorch).

Simulators and Schedulers

mille‑feuille implements abstract base classes to interact with generic executable simulators. These follow a simple design pattern that simulators are expected to follow:

1. External input files are written based on the candidate X values
2. Batches of the (mpiexec'ed) simulator are launched via a Scheduler
3. The simulator writes output files which must be post-processed to extract useful information (P) and the objective function value (Y)
4. (Optionally) A clean up of the Simulator and Schedular output files is performed

Take a look at the examples and the tests to see implementations of Simulators and Schedulers.

Additional Info

CI makes use of cached IntelOneAPI install by scivision

✉️ Contact

Aidan Crilly · ac116@ic.ac.uk

Licensed under MIT © 2025