Interpretable Machine Learning (XGBoost + SHAP) for global fits in high energy physics and cosmology. This repository gathers the code, Jupyter notebooks and data behind the three applications presented in the accompanying paper, where an XGBoost surrogate of the (generally non-Gaussian) log-likelihood accelerates parameter-space exploration by several orders of magnitude while SHAP values keep the inference physically interpretable.
Each subproject lives in its own repository with isolated dependencies and provides the full pipeline used in the paper: training-set construction, XGBoost surrogate training, SHAP analysis and MCMC sampling of the emulated posterior.
| Project | Application |
|---|---|
| BMesonsML | Flavour anomalies in semileptonic B-meson decays: surrogate of the SMEFT19 global likelihood in Scenario III (C1, C3, βq), likelihood maps, SHAP ranking and observable correlations |
| AlpsML | Axion-like particles as a solution to the Belle II B⁺→K⁺νν̄ excess: UV models with PQ charges, alpaca likelihood, two-stage classifier + regressor surrogate and MCMC posterior |
| CosmoML | Cosmology: ΛCDM and w₀wₐCDM fits to SNe Ia (Pantheon+, DES-SN5YR) and BAO (DESI DR2) with a compressed Planck CMB prior; doubles as a controlled validation against the exact likelihood |
All subprojects share the same ML pattern:
- XGBoost trained as a surrogate of the χ² (or a transform of it chosen per application: direct log-likelihood, sigmoid classifier + regressor, or a shifted-log₁₀ target).
- SHAP to rank parameter importance and expose degeneracies, keeping the emulated fit interpretable.
- MCMC (Metropolis-Hastings or
emcee) run on the emulated likelihood to map posteriors and confidence regions. - Jupyter notebooks as the interface, with reusable logic in an importable package per subproject where needed.
Domain-specific dependencies (astropy, alpaca, flavio, wilson, smelli,
…) are kept isolated in each repo's requirements.txt / pyproject.toml.