MadNIS at NLO
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.
Kai: We're looking at the paper: "MadNIS at NLO".
Mira: Here's the gist of the paper. Precise and scalable event generation is the key theme in theoretical particle physics as the upcoming High-Luminosity LHC (HL-LHC) will push complexity and luminosity to unprecedented levels > The unified NLO implementation of ultrafast amplitude surrogates and neural importance sampling is the natural next step in ML-enhanced event generation.
Kai: FKS Subtraction Formalism The NLO cross section is written as a sum of Born-like, virtual, and real emission contributions over the Born-like phase space and the (n+1)-particle phase space > The real emission phase space is regularized by a suitable subtraction scheme where the (n+1)-particle subtraction term dσS has local divergences similar to dσR and its integral dσI cancels the corresponding divergence in dσV.
Mira: The FKS subtraction scheme splits the real emission phase space into FKS sectors for all possible pairs of particles that can introduce soft, collinear or soft–collinear singularities in the real matrix element.
Kai: Amplitude Surrogate Models Learned amplitude surrogates are introduced as the first key ingredient for ultra-fast NLO calculations.
Paper summary: Kai: Looking at the paper "MadNIS at NLO", the core message is that combining fast amplitude surrogates with neural importance sampling creates a coherent machine learning framework for subtraction-based NLO calculations <ref:2603.22407#pg1>.
Mira: The authors found that learning the virtual to Born ratio performed best without including the integrated subtraction contribution, even though incorporating it is straightforward <ref:2603.22407#pg1>.
Lev: So for real emission, they stick with FKS subtraction and train sector conditioned surrogates of those regularized integrands away from divergences <ref:2603.22407#pg1>, meaning the precision from subtracting surrogates on the real emission side will actually be seriously degraded <ref:2603.22407#pg1>.
Kai: The figure of merit they focus on is acceleration at a given precision, where they found speed gains of a factor one hundred ten for NLO three jet predictions and a factor five hundred seventy for NLO four jet predictions <ref:2603.22407#pg1>.
Mira: The overall implication is that this comprehensive surrogate approach makes the entire workflow compatible with GPU parallelization, which is really important when you think about simulating things at the HL-LHC <ref:2603.22407#pg1>.
Lev: It’s a method for accelerating those expensive simulations, but we need to remember that this is based on specific assumptions about how well those learned surrogates map to the actual matrix elements <ref:2603.22407#pg1>.
Conclusion: Kai: So we've been looking at this paper about MadNIS at NLO, and basically, they've put together a method to speed up those very hard calculations in particle physics.
Mira: Right, it’s about taking these complex next-to-leading order calculations and making them much faster by using machine learning for the important parts of the math.
Kai: The authors are focused on combining amplitude surrogates with neural importance sampling, which is a pretty smart way to tackle this kind of problem.
Mira: They show that when you use these things together, you get really fast results for NLO three jet predictions and even faster ones for four jet predictions.
Kai: But the key thing they're pointing out is that this isn't just about getting a few more calculations done; it’s about how much faster we can get them at the same level of accuracy.
Mira: They found that the virtual to Born ratio learning works really well without needing to include those integrated subtraction terms, which is interesting because incorporating them is actually pretty straightforward too.
Kai: It means for real emission, they're using FKS partitioning and training sector conditioned surrogates instead of trying to subtract the full contribution, which means the precision on that part takes a hit.
Mira: So the big picture here is that this whole setup can be parallelized on GPUs, which is crucial when you’re dealing with the kind of huge datasets we expect from things like a future high luminosity collider.
Kai: It shows how ML frameworks can work alongside traditional subtraction methods to tackle these precision physics problems.
Mira: The next thing we need to look at is how this actually translates into real-world hardware for error correction, which is where the engineering really gets interesting.
Institut für Theoretische Physik, University of Heidelberg, Germany · DAMTP, University of Cambridge, United Kingdom · §CP3, Universite catholique de Louvain, Louvain-la-Neuve, Belgium · Interdisciplinary Center for Scientific Computing (IWR), University of Heidelberg
hep-ph
Submitted: 2026-03-23
Updated: 2026-11-11
Code: https://github.com/madgraph-ml/madnis-nlo
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 100/100
The gist: Precise and scalable event generation is the key theme in theoretical particle physics as the upcoming High-Luminosity LHC (HL-LHC) will push complexity and luminosity to unprecedented levels > The…
Key concepts
- Amplitude Surrogate Models
- These are machine learning models trained to quickly estimate the amplitude of a physical process. Instead of calculating the full, complex mathematical expression every time, the network learns to map input data (like particle momenta) directly to the amplitude value. This drastically speeds up calculations.
- FKS Subtraction Formalism
- This is a method used in NLO calculations to handle infinities arising from real particle emissions. It splits the real emission phase space into smaller sectors based on which particles are close together, allowing for systematic subtraction of these singularities to get a finite result.
- Neural Importance Sampling (MADNIS)
- This technique uses neural networks to guide Monte Carlo integration. By using the FKS partitioning and multi-channel mappings, the model learns how to sample the phase space efficiently, focusing computational effort where it is most needed, leading to faster and more accurate integration.
Terminology
Summary
Precise and scalable event generation is the key theme in theoretical particle physics as the upcoming High-Luminosity LHC (HL-LHC) will push complexity and luminosity to unprecedented levels > The unified NLO implementation of ultrafast amplitude surrogates and neural importance sampling is the natural next step in ML-enhanced event generation. p. 1 SciPost Physics Submission
1 Introduction
Precise and scalable event generation is the key theme in theoretical particle physics [1],
as the upcoming High-Luminosity LHC (HL-LHC) will push complexity and luminosity to
unprecedented levels. Event generators such as PYTHIA [2], SHERPA [3], HERWIG [4], and
MadGraph [5–7], specifically MG5AMC [8,9], provide the backbone of the first-principles simulation chain, combining perturbative QCD calculations with parton showers and hadronization.Where the paper says it
FKS Subtraction Formalism
The NLO cross section is written as a sum of Born-like, virtual, and real emission contributions over the Born-like phase space and the (n+1)-particle phase space > The real emission phase space is regularized by a suitable subtraction scheme where the (n+1)-particle subtraction term dσS has local divergences similar to dσR and its integral dσI cancels the corresponding divergence in dσV. The FKS subtraction scheme splits the real emission phase space into FKS sectors for all possible pairs of particles that can introduce soft, collinear or soft–collinear singularities in the real matrix element. p. 2 Beyond the divergence, the form of the subtraction term dσ
S
varies. We employ the FKS subtraction scheme [77, 78], which splits the real emission phase space into FKS sectors for all
possible pairs of particles that can introduce soft, collinear or soft–collinear singularities in the
real matrix element.Where the paper says it
Amplitude Surrogate Models
Learned amplitude surrogates are introduced as the first key ingredient for ultra-fast NLO calculations. These surrogates encode individual contributions or the combined Born-like amplitude by training a regression network on phase-space functions. p. 5 A network
surrogate can encode individual contributions or the combined Born-like amplitude. We generate a set of external momenta with MADNIS and train a regression network to learn the
phase-space functions
Partial sum A
BV = A
B + A
V
Ratio V/B R
V/B =
AV
AB
Total sum A
BVI = A
B + A
V + A
I
Ratio (VI)/B R
(VI)/B =
AV + AI
AB
. As input, we combine the set of final-state 4-momenta and the log-invariants
y
B = (Φn
, log s
B
kl) with s
B
kl = pk
· pl
for k ̸= l . (20)
Over this phase space, we learn the logarithmic amplitude or amplitude-ratio surrogates
fθ
( y
B
) ≈ f ( y
B
) with f ∈ {logA,R} .Where the paper says it
Where the paper says it
Neural Importance Sampling
Neural importance sampling is extended to NLO by adapting the multi-channel formalism to include the FKS partitioning of the real emission phase space. The integral is rewritten using a complete mapping (xB, xrad) for each channel k and FKS sector i j. p. 0 For real emission, we stick to the standard FKS subtraction and train sector-conditioned surrogates
of the regularized integrands away from divergences. MADNIS then uses multi-channel
mappings and FKS sectors as conditions. For real emission, we stick to the standard FKS subtraction and train sector-conditioned surrogates
of the regularized integrands away from divergences. MADNIS then uses multi-channel
mappings and FKS sectors as conditions.Where the paper says it
Where the paper says it
Where the gains come from
The acceleration through ML-surrogates is shown in Fig. 12 where the average integrand evaluation time versus relative variance of the integral. p. 4 Physical predictions combining Born-like and real emission contributions are formally independent of these parameters, but they can have an impact on the efficiency of the
numerical integration. Indeed, the localization of the cancellations affects the variance of the
Monte Carlo integral and influences the fraction and distribution of negative event weights. Supplementing these various surrogates, neural importance sampling [63–67] has been successfully applied at leading order (LO) using MADNIS [68–70] or its SHERPA counterpart [71–73]. Normalizing-flow samplers have also been used at NLO [71] and NNLO accuracy in multi-jet
final states [74]. (16)
Illustrative Feynman diagrams for the Born, virtual, and real-emission contributions to the 4-jet
process are shown in Fig. Using the 3-jet case for illustration, with analogous considerations
applying to the 4-jet case, we highlight some aspects of the singularity structure:
• For e+e
− → uu gg, there exist five collinear configurations, each gluon can become collinear ¯
to the quark or antiquark, or the two gluons can form a collinear pair.Where the paper says it
Where the paper says it
Where the paper says it
Performance and Optimization
The optimization of the subtraction threshold is performed by balancing integrand variance against potential speed gains from the surrogate in three ways: collinear threshold δ = λ ξcut = 0.5, soft threshold δ = 1. p. 0 2 Neural importance sampling 15
5 Performance 15
5. 1 Optimized subtraction threshold 15
5. The phase-space construction is discussed in more detail in Sec. The fully subtracted real-emission contribution can then be written as
σn+1 =
X
i j
Z
dΦ
(i j)
n+1
1
2s
AR-S
i j (Φ
(i j)
n+1
)
Nn+1
≡
X
i j
Z
dΦ
(i j)
n+1
f
i j
n+1
(Φ
(i j)
n+1
) , (12)
with
A
R−S
i j (Φ
(i j)
n+1
) = 1
ξ
2
i
(1 − yi j)
Σi j(Φn
,ξi
, yi j,ϕi
)
−
∂ Φ
coll
n+1
∂ Φ
hard
n+1
Σi j(Φn
,ξi
, 1,ϕi
)Θ( yi j − 1 + δ)
−
∂ Φ
soft
n+1
∂ Φ
hard
n+1
Σi j(Φn
, 0, yi j,ϕi
)Θ(ξcut − ξi
)
+
∂ Φ
soft–coll
n+1
∂ Φ
hard
n+1
Σi j(Φn
, 0, 1,ϕi
)Θ( yi j − 1 + δ)Θ(ξcut − ξi
)
. Supplementing these various surrogates, neural importance sampling [63–67] has been successfully applied at leading order (LO) using MADNIS [68–70] or its SHERPA counterpart [71–73]. Normalizing-flow samplers have also been used at NLO [71] and NNLO accuracy in multi-jet
final states [74]. Supplementing these various surrogates, neural importance sampling [63–67] has been successfully applied at leading order (LO) using MADNIS [68–70] or its SHERPA counterpart [71–73]. Normalizing-flow samplers have also been used at NLO [71] and NNLO accuracy in multi-jet
final states [74]. SciPost Physics Submission
ratio with the Born amplitude leads to a competitive accuracy. In terms of deviations from
the actual amplitudes, we find that for the 3-jet process there are essentially no phase-space
points with deviations larger than one per-mille, and for the 4-jet process there are hardly
any phase-space points with deviations above a percent.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Outlook
The paper presents a coherent ML framework for subtraction-based NLO calculations combining amplitude surrogates with neural importance sampling. The authors found that learning the virtual-to-Born ratio performed best without including the integrated subtraction contribution, but incorporating it is straightforward. p. 1 Building on this structure, we
provide an NLO version of the MADNIS framework. For virtual corrections, we find that learning the ratio of the subtracted virtual correction to the Born matrix element provides the best
balance between speed and precision. SciPost Physics Submission
The terms in brackets consist of the locally regularized real-emission contribution together
with its collinear, soft, and soft–collinear subtraction terms, weighted by its corresponding
phase-space factor. The parameters ξcut and δ define the regions in which the subtractions
are active. Further details on the network hyperparameters are given in Tab. 10−19 10−16 10−13 10−10 10−7 10−4 10−1
Σ
101
102
103
104
105
106
Counts
e+
e− → uug (@1 TeV) ¯
Learned Σ
Truth
10−22 10−18 10−14 10−10 10−6 10−2
Σ
101
102
103
104
105
106
Counts
e+
e− → uugg (@1 TeV) ¯
Learned Σ
Truth
Figure 5: Learned real emission amplitudes for the NLO corrections to 3-jet (left) and
4-jet (right) production. MADNIS then uses multi-channel
mappings and FKS sectors as conditions. We validate our approach for electron-positron
scattering to three and four jets and find significant speed-ups and variance reduction
in the integration.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Code Availability
The code used in this work is publicly available on GitHub as part of the ML for MadGraph organization in the repository https://github.com/madgraph-ml/madnis-nlo. p. 0 SciPost Physics Submission
MadNIS at NLO
Giovanni De Crescenzo1
, Javier Mariño Villadamigo1
,
Nina Elmer2
, Theo Heimel3
, Tilman Plehn1,4, Ramon Winterhalder5
, Marco Zaro5
1 Institut für Theoretische Physik, Universität Heidelberg, Germany
2 DAMTP, University of Cambridge, Cambridge, United Kingdom
3 CP3, Université catholique de Louvain, Louvain-la-Neuve, Belgium
4 Interdisciplinary Center for Scientific Computing (IWR), Universität Heidelberg, Germany
5 TIFLab, Università degli Studi di Milano & INFN Sezione di Milano, Italy
April 9, 2026
Abstract
We combine fast amplitude surrogates with neural importance sampling to accelerate
NLO calculations. For virtual corrections, a learned ratio to the Born matrix element with
calibrated uncertainties guarantees reliable precision across phase space. Next-to-leading order (NLO) and even higher order predictions are essential for precision
LHC physics, but their complex phase-space integrations and repeated evaluations of expensive
matrix elements constitute a major computational bottleneck. The precision and simulation
statistics required by the HL-LHC implies a rapidly growing computational cost and motivates
the use of modern machine learning (ML) [10, 11] to accelerate all components of the simulation pipeline [12] and the simulation workflow [13].Where the paper says it
Where the paper says it
Hyperparameters
Table 5 lists the hyperparameters for MADNIS integration runs, including batch size and number of MADNIS iterations. Table 3 details the hyperparameters for real emission surrogates. p. 8 Further details on the network hyperparameters are given in Tab. 10−19 10−16 10−13 10−10 10−7 10−4 10−1
Σ
101
102
103
104
105
106
Counts
e+
e− → uug (@1 TeV) ¯
Learned Σ
Truth
10−22 10−18 10−14 10−10 10−6 10−2
Σ
101
102
103
104
105
106
Counts
e+
e− → uugg (@1 TeV) ¯
Learned Σ
Truth
Figure 5: Learned real emission amplitudes for the NLO corrections to 3-jet (left) and
4-jet (right) production. The network architecture is again a simple MLP with hyperparameters listed in
Tab. The discrete FKS sector index is provided through a look-up table and a linear layer.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Virtual correction surrogates can be improved by learning
the ratio of the subtracted virtual correction to the Born matrix element with calibrated uncertainties,
whichguarantees reliable precision across phase space.
-
Real emission surrogates can be made more robust by restricting them to
phase space regions without subtraction
because using them in subtraction schemes may lead to a significant decrease in accuracy due toextremely sensitive cancellations.
-
The MADNIS framework can be extended to include the FKS partitioning of the real emission phase space, allowing it to use
multi-channel neural importance sampling combined with FKS sectors
as additional discrete degrees of freedom. -
The AI system can perform NLO calculations by combining learned amplitude surrogates and neural importance sampling, achieving a
factor 110 for NLO 3-jet predictions and a factor 570 for NLO 4-jet predictions
speedup over VEGAS. -
The system can be optimized to choose the best subtraction thresholds by balancing "the integrand variance against the potential speed gains from the surrogate in three ways: collinear threshold delta = lambda = 0.5, soft threshold delta = lambda = 1.0, combined thresholds delta = lambda xcut = lambda with lambda in [1e-4, 1]."
Abstract
We combine fast amplitude surrogates with neural importance sampling to accelerate NLO calculations. For virtual corrections, a learned ratio to the Born matrix element with calibrated uncertainties guarantees reliable precision across phase space. For real emission, we stick to the standard FKS subtraction and train sector-conditioned surrogates of the regularized integrands away from divergences. MadNIS then uses multi-channel mappings and FKS sectors as conditions. We validate our approach for electron-positron scattering to three and four jets and find significant speed-ups and variance reduction in the integration.
Sources
- Event Generators for High-Energy Physics Experiments
- A comprehensive guide to the physics and usage of PYTHIA 8.3
- Event generation with Sherpa 3
- The Physics of Herwig 7
- MadEvent: Automatic Event Generation with MadGraph
- MadGraph/MadEvent v4: The New Web Generation
- MadGraph 5 : Going Beyond
- The automated computation of tree-level and next-to-leading order differential cross sections, and their matching to parton shower simulations
- The automation of next-to-leading order electroweak calculations
- Modern Machine Learning for LHC Physicists
- Modern Machine Learning and Particle Physics Phenomenology at the LHC
- Machine Learning and LHC Event Generation
- MadAgents
- (Machine) Learning amplitudes for faster event generation
- Using neural networks for efficient evaluation of high multiplicity scattering amplitudes
- Optimising simulations for diphoton production at hadron colliders using amplitude neural networks
- A factorisation-aware Matrix element emulator
- Accelerating Monte Carlo event generation -- rejection sampling using neural network event-weight estimates
- Targeting Multi-Loop Integrals with Neural Networks
- Unweighting multijet event generation using factorisation-aware neural networks
Related papers
- The NANOGrav 15-year data set: A Limit on Continuous Spin Gravitons
- DIS-secting the universe with high-energy neutrinos
- Crowdsourcing Gravitational Waves from Superradiant Axions
- Classification of g-modes for neutron stars with a strong transition: Novel universal relation including slow stable hybrid stars
- Scalar axion field of toroidal electromagnetic pulses
- A framework for matrix-element-corrected QED antenna showers in hadron decays