An open-source machine learning framework for global analyses of parton distributions

Abstract

We present the software framework underlying the NNPDF4.0 global determination of parton distribution functions (PDFs). The code is released under an open source licence and is accompanied by extensive documentation and examples. The code base is composed by a PDF fitting package, tools to handle experimental data and to efficiently compare it to theoretical predictions, and a versatile analysis framework. In addition to ensuring the reproducibility of the NNPDF4.0 (and subsequent) determination, the public release of the NNPDF fitting framework enables a number of phenomenological applications and the production of PDF fits under user-defined data and theory assumptions.

Event Generators for High-Energy Physics Experiments

“…One successful application of machine learning to particle physics phenomenology was the development of the NNPDF parton densities [413][414][415]. It is expected that modern machine learning techniques can improve LHC simulations in other aspects as well [416].…”

Parton distributions need representative sampling

“…These PDF solutions are linear superpositions of the NNPDF4.0 Hessian replicas available in the LHAPDF library [42]. The χ 2 values are computed using the published NNPDF4.0 code [24,25] in accord with the convention adopted in the comparison tables in Sec. 5.1 of the NNPDF4.0 publication.…”

“…We generate LHAPDF6 tables for the sample PDF replicas using the mcgen program [21][22][23] and the LHAPDF tables of the NNPDF4.0 NNLO Hessian ensemble as the input. The total χ 2 of the NNPDF4.0 analysis was evaluated using the public code released by NNPDF [24,25], without refitting. Specifically, the χ 2 is computed by the perreplica chi2 table function of program validphys included in the NNPDF code.…”

“…II D and present an example application in Sec. III, where we investigate the PDF uncertainty on the LHC benchmark cross sections using the NNPDF4.0 error sets and publicly available mcgen [21][22][23] and NNPDF [24,25] fitting codes. The hopscotch scan+sampling technique introduced there suggests that the PDF uncertainty on key LHC cross sections at 13 TeV is larger than the nominal uncertainty obtained with the published NNPDF4.0 error sets.…”

Machine Learning and LHC Event Generation

“…The way to approach this problem as a machine learning challenge was first suggested long ago [47]: the basic idea is to deliver a Monte Carlo ensemble of machine learning models, such as neural networks, that provide the desired representation of a probability of probabilities. The successful implementation of this idea has led to the NNPDF family of proton PDF determinations [46,[48][49][50] as well as to variants in the context of polarised PDF [51] and nuclear PDF [52,53] global analyses. The current implementation frontier, which has led to the recent NNPDF4.0 determination, involves a suite of contemporary machine learning methods and tools, specifically cross-validation to avoid overtraining, hyperoptimization [54] combined with K-folding for the automatic selection of the methodology, feature scaling of the input for the optimization of the neural networks used as basic underlying model [55], and GAN-enhanced compression for final efficient delivery [56,57].…”