[論文レビュー] The MadNIS Reloaded

Theo Heimel, Nathan Huetsch+4 more|arXiv (Cornell University)|Nov 2, 2023

Particle physics theoretical and experimental studies参考文献 111被引用数 7

ひとこと要約

The paper implements and enhances the MadNIS method within MadGraph MG5aMC, introducing new ML-driven components and training strategies that substantially improve phase-space sampling and unweighting efficiency for challenging LHC processes.

ABSTRACT

In pursuit of precise and fast theory predictions for the LHC, we present an implementation of the MadNIS method in the MadGraph event generator. A series of improvements in MadNIS further enhance its efficiency and speed. We validate this implementation for realistic partonic processes and find significant gains from using modern machine learning in event generators.

研究の動機と目的

• Provide an end-to-end ML-augmented MadNIS implementation inside MG5aMC for fast, precise LHC predictions.
• Improve phase-space sampling and variance reduction via neural channel weights and learned mappings.
• Demonstrate performance gains (unweighting efficiency and sampling precision) across representative high-multiplicity processes.
• Explore training strategies (online/buffered, stratified sampling, channel dropping) to stabilize and speed up learning.
• Analyze how learned channel weights reflect amplitude patterns and diagram group importance.

提案手法

• Encode local multi-channel weights and phase-space mappings as neural networks (CWnet for channel weights and INN/RQS for phase-space mappings).
• Use a softmax-based normalization to enforce channel weights and initialize with physics-inspired priors (MG and P-based priors).
• Replace Vegas with an invertible neural network (INN) for latent-space phase-space mapping, ensuring fast forward/inverse evaluations.
• Train using a variance-minimizing loss (MadNIS loss) that accounts for per-channel variances and optimal sample allocation across channels.
• Employ online and buffered training with stratified sampling and channel dropping to stabilize and accelerate learning and reduce channel count.
• Initialize INN mappings with a Vegas-based pretraining to leverage factorization of the integrand.

実験結果

リサーチクエスチョン

• RQ1 Can MadNIS be effectively implemented in MG5aMC to yield tangible speedups for realistic LHC processes?
• RQ2 How do neural channel weights and learned phase-space mappings improve unweighting efficiency and integration variance compared to standard Vegas?
• RQ3 What is the impact of training strategies (online/buffered, stratified sampling, channel dropping) on stability and performance across high-multiplicity final states?
• RQ4 How do learned channel weights reflect underlying amplitude patterns and diagram-group importance in different processes?
• RQ5 How does MadNIS scale with increasing jet multiplicity and complex final states?

主な発見

Process	# diagrams	# channels	# channel groups	# active channels
Triple-W	17	16	8	2 … 4
VBS	51	30	15	4 … 6
W+jets (gg→W+ d u)	8	8	4	6
W+jets (gg→W+ d u g)	50	48	24	12 … 16
W+jets (gg→W+ d u g g)	428	384	108	28 … 51
t tbar +jets (gg→t tbar + g)	16	15	9	4 … 6
t tbar +jets (gg→t tbar + g g)	123	105	35	12
t tbar +jets (gg→t tbar + g g g)	1240	945	119	60 … 72

• MadNIS implementations in MG5aMC yield significant gains in unweighting efficiency for challenging processes (e.g., up to an order of magnitude improvement noted in abstract).
• For VBS, unweighting efficiency reaches about 20%, a factor of over ten improvement versus the standard method.
• Stratified training combined with trained channel weights yields large performance gains, up to a factor of 15 for VBS and notable gains for other processes.
• Channel dropping stabilizes training in high-channel-count setups by focusing on the most impactful channels/groups.
• Learned channel weights reveal that MadNIS concentrates sampling on a small subset of symmetry-related channel groups, sometimes dominating the integral.
• Across W+jets and ttbar+jets scenarios, MadNIS maintains substantial gains as jet multiplicity increases, though the gain diminishes with very high multiplicities (notably ttbar+3 jets).