DLScanner: A parameter space scanner package assisted by deep learning methods

TL;DR

DLScanner tackles the challenge of efficiently exploring high-dimensional BSM parameter spaces by fusing similarity learning with VEGAS adaptive sampling in an iterative, feedback-driven loop. It supports both DL regressors and classifiers (including MLP and SL) to map sampling spaces to target regions, with VEGAS concentrating samples where they matter most. The MSSM case demonstrates faster convergence when using SL or MLP classifiers with VEGAS compared to random sampling, and highlights the robustness and generality of the approach across tools such as SPheno and micrOMEGAs. This framework offers a practical, open-source path to accelerate parameter-space investigations with large, expensive-to-evaluate observables, enabling more efficient and comprehensive explorations of BSM theories.

Abstract

In this paper, we introduce a scanner package enhanced by deep learning (DL) techniques. The proposed package addresses two significant challenges associated with previously developed DL-based methods: slow convergence in high-dimensional scans and the limited generalization of the DL network when mapping random points to the target space. To tackle the first issue, we utilize a similarity learning network that maps sampled points into a representation space. In this space, in-target points are grouped together while out-target points are effectively pushed apart. This approach enhances the scan convergence by refining the representation of sampled points. The second challenge is mitigated by integrating a dynamic sampling strategy. Specifically, we employ a VEGAS mapping to adaptively suggest new points for the DL network while also improving the mapping when more points are collected. Our proposed framework demonstrates substantial gains in both performance and efficiency compared to other scanning methods.

Figures (6)

• Figure 1: Charts for the iterative processes used for the regressor (left) and the classifier (right). Black arrows indicate the main predict-train loop, green arrows indicate places where random input is required, orange arrows mark the parts where we collect the output dataset and blue arrows highlight where VEGAS mapping is trained and used to suggest new points.
• Figure 2: Schematic representation of similarity learning within the scanning loop. During training, the network maps sampled points that satisfy the constraints to a different region of the representation space than those that do not. Once the representation space is structured, it can then be used to predict points that are very close to the target region.
• Figure 3: Collected samples for the function $O_\text{3d}$ as described in the text. In the left, the total of samples collected in and out of target are displayed (90 000 points, opacity of each point is 0.2). In the right, only the samples collected in target are displayed (17 000 points, opacity of each point is 0.5). The points in the left correspond to the points that were selected by the network and ultimately passed to $O_\text{3d}$.
• Figure 4: Progression of reduction of samples in the process used here. The leftmost panel corresponds to the distribution of points generated when using VEGAS map for a total of $10^6$ points (opacity of each point is 0.1). The panel in the middle corresponds the filter using the network to predict which points are either in- or out-target and selecting only $10^4$ points (opacity of each point is 0.2). The rightmost panel corresponds to the final points that would be collected as in-target during the step, in this case 5430 points (opacity of each point is 0.2).
• Figure 5: Number of accumulated valid points in terms of number of iterations for different scanning methods. Solid lines represent the VEGAS sampled scan while dashed lines represent the random sampling.