Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MadSpace -- Event Generation for the Era of GPUs and ML

Theo Heimel, Olivier Mattelaer, Ramon Winterhalder

TL;DR

MadSpace introduces a GPU- and ML-friendly, compute-graph-based framework for end-to-end LO parton-level event generation. It unifies phase-space construction, adaptive sampling (VEGAS and MadNIS with normalizing flows), PDF interpolation on-device, phase-space cuts, weighted histograms, and on-device unweighting, with a Python interface via DLPack. A Unified Matrix Element Interface (UMAMI) enables batched, device-resident matrix-element calls without process-specific code generation, paving the way for end-to-end GPU workflows and future NLO support. Validation against MG5aMC demonstrates correct phase-space volumes, faithful differential distributions, and substantial throughput improvements, especially on GPUs, aided by binary intermediate formats and in-memory channel orchestration; the work outlines paths toward neural samplers and differentiable event generation as future enhancements.

Abstract

MadSpace is a new modular phase-space and event-generation library written in C++ with native GPU support via CUDA and HIP. It provides a unified compute-graph-based framework for phase-space construction, adaptive and neural importance sampling, and event unweighting. It includes a wide range of mappings, from the standard MadGraph multi-channel phase space to optimized normalizing flows with analytic inverse transformations. All components operate on batches of events and support end-to-end on-device workflows. A high-level Python interface enables seamless integration with machine-learning libraries such as PyTorch.

MadSpace -- Event Generation for the Era of GPUs and ML

TL;DR

MadSpace introduces a GPU- and ML-friendly, compute-graph-based framework for end-to-end LO parton-level event generation. It unifies phase-space construction, adaptive sampling (VEGAS and MadNIS with normalizing flows), PDF interpolation on-device, phase-space cuts, weighted histograms, and on-device unweighting, with a Python interface via DLPack. A Unified Matrix Element Interface (UMAMI) enables batched, device-resident matrix-element calls without process-specific code generation, paving the way for end-to-end GPU workflows and future NLO support. Validation against MG5aMC demonstrates correct phase-space volumes, faithful differential distributions, and substantial throughput improvements, especially on GPUs, aided by binary intermediate formats and in-memory channel orchestration; the work outlines paths toward neural samplers and differentiable event generation as future enhancements.

Abstract

MadSpace is a new modular phase-space and event-generation library written in C++ with native GPU support via CUDA and HIP. It provides a unified compute-graph-based framework for phase-space construction, adaptive and neural importance sampling, and event unweighting. It includes a wide range of mappings, from the standard MadGraph multi-channel phase space to optimized normalizing flows with analytic inverse transformations. All components operate on batches of events and support end-to-end on-device workflows. A high-level Python interface enables seamless integration with machine-learning libraries such as PyTorch.
Paper Structure (18 sections, 1 equation, 8 figures, 1 table)

This paper contains 18 sections, 1 equation, 8 figures, 1 table.

Table of Contents

  1. Features
  2. Phase space
  3. Adaptive sampling
  4. PDF interpolation
  5. Phase-space cuts
  6. Weighted histograms
  7. Renormalization and factorization scale choices
  8. Python interface
  9. UMAMI -- A unified matrix element interface
  10. Unweighting and file output
  11. Event generation
  12. Binary event format
  13. Combining channels and final event output
  14. Validation and performance
  15. Phase-space volume and inverse mappings
  16. ...and 3 more sections

Figures (8)

  • Figure 4: Illustration of the synchronous CPU graph execution mode (left) and asynchronous CPU and GPU execution modes (right).
  • Figure 5: Compute graph generated by MadSpace for an example diagram with both $t$-channel (%24) and $s$-channel (%26) propagators.
  • Figure 6: Illustration of the full event-generation workflow.
  • Figure 7: From left to right: A pure $t$-channel, mixed $t$- and $s$-channel, and pure $s$-channel topology for a generic $2\to4$ process.
  • Figure 8: Left: computed phase-space volume for different mappings relative to the analytical result for a massless $2\to4$ process. Points were sampled until the target precision of $10^{-4}$ was reached. Right: histogram of the relative deviation between the random inputs and recovered random outputs when the forward and inverse mappings are evaluated subsequently for 1M points flattened over the 10 random dimensions.
  • ...and 3 more figures