Arbitrary Data Walkthrough

Source files:

• examples/arbitrary-data/

Goal

This notebook-first example shows how to use BFF when simulation results and reference targets already exist outside the built-in trajectory-analysis workflow. It calibrates two water-like Lennard-Jones parameters against realistic synthetic density, enthalpy-of-vaporization, and diffusion data.

No GROMACS installation is required. The fitting and learning cells use CUDA by default.

Run

Install the optional notebook tools once:

pip install "bfflearn[notebook]"

cd examples/arbitrary-data
jupyter lab

Open arbitrary-data.ipynb and execute it from top to bottom. The notebook demonstrates the complete data-facing workflow:

1. load user-provided whitespace-delimited tables;
2. construct and write one QoIDataset per observable;
3. fit local Gaussian-process surrogate models;
4. build a constrained LearningProblem;
5. sample the posterior and write standard and QoI-attributed marginal plots.

The generated qoi-*.pt datasets can also be passed to bff fit. All notebook-generated files are written under generated/. The plot generated/qoi-marginals.pdf colors each posterior marginal by the local contribution from density, enthalpy of vaporization, and diffusion.

Input Tables

• raw-data/simulation-results.dat contains one externally evaluated parameter set per row.
• raw-data/experimental-targets.dat contains the target value and uncertainty for each observable.
• The inline Specs dictionary in the notebook defines the parameter bounds used during posterior learning.

The committed .dat values are illustrative synthetic data. Replace them with your own tables and update the notebook mappings for a new application.