A coverage study of the CMSSM based on ATLAS sensitivity using fast neural networks techniques

Michael Bridges, K. Cranmer, Farhan Feroz, M. P. Hobson, Roberto Ruiz de Austri, Roberto TrottaJournal of High Energy Physics 2011 · DOI · arXiv

Abstract

This study addresses the computational bottleneck in performing coverage tests for the Constrained Minimal Supersymmetric Standard Model (CMSSM). By implementing Artificial Neural Networks (ANNs) to emulate complex Renormalization Group Equation (RGE) solvers, the researchers achieved a speedup of approximately 30,000 times. This acceleration enabled a comprehensive coverage study using $10^4$ pseudo-experiments based on ATLAS sensitivity. The findings reveal that both Bayesian and Profile Likelihood intervals exhibit significant over-coverage, largely due to the physical boundaries of the model's parameter space.

Introduction

In high-energy physics, the search for new phenomena often relies on the CMSSM. Validating the statistical intervals (confidence or credible intervals) derived from experimental data is crucial. A "coverage study" verifies if a 95% interval actually contains the true value in 95% of cases. However, such studies are traditionally computationally prohibitive, requiring millions of likelihood evaluations per pseudo-experiment. For the CMSSM, a full study would take over 1,100 CPU-years using standard tools like SoftSusy.

Methodology

To overcome this hurdle, the authors developed a machine-learning pipeline. A 3-layer Multilayer Perceptron (MLP) was trained to map high-energy CMSSM parameters to the low-energy sparticle mass spectrum. Additionally, a classification network was used to identify unphysical regions, such as those failing electroweak symmetry breaking, allowing the sampler to skip invalid points.

The study utilized the ATLAS "SU3" benchmark point as a reference. The parameter space explored included both CMSSM fundamental parameters and Standard Model (SM) nuisance parameters, as detailed in the table below.

Table of CMSSM parameters and SM nuisance parameters defining the search space.

Sampling was performed using MultiNest and Markov Chain Monte Carlo (MCMC) algorithms. The ANN-based approach reduced the total computation time from centuries to approximately 20,000 CPU-minutes.

Results and Key Insights

The ANN emulation proved highly accurate, matching dedicated physics codes within a few percent. The primary result of the coverage study is the observation of "over-coverage." For instance, intervals designed for 68% coverage often reached over 80%.

3D visualization of the CMSSM parameter space surfaces generated by the neural network emulator.

This phenomenon is attributed to the "Boundary Effect." When a best-fit point falls into an unphysical region, the model constraints force the statistical interval to stay within physical bounds, effectively widening the interval and making it more conservative. While this reduces the risk of false discoveries, it may also mask the true sensitivity of the experiment.

Conclusion

The integration of ANNs into the particle physics statistical pipeline offers a transformative speedup for complex model validation. Future work could expand these techniques to global scans of the CMSSM or other computationally expensive observables like dark matter relic density. This methodology ensures that the "trust but verify" principle can be applied even to the most complex theories in modern physics.

Summary updated

BibTeX
@article{bridges2011coverage,
  author = {Michael Bridges and K. Cranmer and Farhan Feroz and M. P. Hobson and Roberto Ruiz de Austri and Roberto Trotta},
  title = {{A coverage study of the CMSSM based on ATLAS sensitivity using fast neural networks techniques}},
  journal = {Journal of High Energy Physics},
  year = {2011},
  doi = {10.1007/jhep03(2011)012},
}
Figure