The full documtation can be found here: https://toolkit-for-simulation-based-inference.readthedocs.io
We will use pixi to setup the environment for the workflow. The specifications are defined in the pixi.toml file. If pixi is not installed on your machine follow the instructions in pixi seutp guide. Then proceed to install the environment with:
pixi install -e nsbi-env-gpu
Currently the environment can only be built on machines with GPU.
A jupyter kernel can then be created by running:
pixi run -e nsbi-env-gpu python -m ipykernel install --user --name nsbi-env-gpu --display-name "Python (pixi: nsbi-env-gpu)"
Simulation-Based Inference (SBI) or Neural Simulation-Based Inference (NSBI) refers to set of statistical techniques that allow statistical inference directly using high-dimensional data. This circumvents the need to build low-dimensional summaries as is traditionally done and which can lose sensitive information.
This toolkit helps facilitate the application of a type of SBI that is scalable for LHC-style analysis with high-dimensional parameter spaces, where the systematic uncertainty modeling is done via certain domain-specific assumptions. This is done via easy-to-use APIs for the various stages in the analysis as well as providing an end-to-end workflow orchestratation pipeline steered via human-readable configuration files.
nsbi-common-utils provides building blocks for SBI analysis tailored to the statistical models typical at the ATLAS and CMS experiments. It implements semi-parametric approach to SBI where the statistical models are built using a combination of non-parametric and parametric methods targeting different parts. The toolkit has a modular structure, and offers APIs for dataset preparation, density-ratio estimation, model building and profiled-likelihood ratio fitting.
The semi-parametric model and workflow is related to the SBI analysis recently published by ATLAS:
- An implementation of neural simulation-based inference for parameter estimation in ATLAS (https://arxiv.org/pdf/2412.01600)
- Measurement of off-shell Higgs boson production in the
$H\to ZZ \to 4\ell$ decay channel using a neural simulation-based inference technique in 13 TeV p-p collisions with the ATLAS detector (https://arxiv.org/pdf/2412.01548)
We demonstrate the usage of nsbi-common-utils applied to a full-scale LHC-style analysis in the examples/. The workflow currently uses the Higgs to tau tau dataset from FAIR universe challenge. More open datasets will be added in the future.
To use the library nsbi_common_utils developed here in general cases outside of this tutorial, do:
python -m pip install --upgrade 'nsbi-common-utils @ git+https://github.com/iris-hep/NSBI-workflow-tutorial.git'Workflow bluprint (tentative):
This work is being supported by the U.S. National Science Foundation (NSF) cooperative agreements OAC-1836650 and PHY-2323298 (IRIS-HEP).
nsbi-common-utils is distributed under the terms of the MIT license.