ML for HEP — The Times of SMEFT

Section B8 – papers tracked 3 years covered Last refreshed –

The Deep Dives · Curated August 29, 2026

Down-Type Jet Identification in Fully Hadronic tt̅ Events

By J. Guo, et al. · arXiv:2608.21982

Fully hadronic tt̅ has the largest branching fraction and a near-maximal spin analyzer in the down-type quark — but nobody uses it for spin correlations because of the jet-assignment nightmare. A two-stage GNN+Transformer pipeline assigns the six-jet system, and a hybrid classifier with a conditional diffusion head resolves the down/up hypotheses inside each W decay.

Why it matters · If jet assignment cracks open the all-hadronic channel, top-quark entanglement and spin-correlation measurements gain their biggest dataset.

Caveat: Phenomenology-level so far; notably, the diffusion score only ranks hypotheses when trained with a margin objective.

Accelerating Optical Photon Simulation in DUNE with Opticks

By I. Parmaksiz et al. · arXiv:2608.27306

First GPU optical-photon simulation at the 10-kiloton scale: Opticks propagates the same photons as GEANT4 through the DUNE far-detector geometry with a 313× speedup over single-threaded GEANT4, validated across all metrics considered.

Why it matters · Optical transport is a canonical simulation bottleneck; two orders of magnitude at full detector scale changes what analyses can afford.

Caveat: Optical photons only, and the integration into production workflows is still in progress.

News in Brief

arXiv:2608.23773 · link — GNN flavour ID for the ESSνSB near detector holds up even when the detector shrinks eightfold — ML as a detector-design cost lever.
arXiv:2608.26224 · link — An honest eight-qubit benchmark: hybrid quantum models vs. classical ML on a CMS-open-data trigger task, under fixed resource budgets.

The Reading Shelf · From the FCC-ee Generators Talk

References cited in Status of event generators for FCC-ee (US FCC Workshop, SLAC, September 2026) that belong to this section — 10 of 53.

The MadNIS reloaded
T. Heimel et al. · SciPost Phys. 17 (2024) 023 · arXiv:2311.01548

MadNIS reloaded: neural importance sampling with multi-channel flows, integrated into MadGraph.
Efficient many-jet event generation with flow matching
E. Bothmann et al. · Phys.Rev.D 114 (2026) 034055 · arXiv:2506.18987

Flow matching for many-jet phase space: the unweighting-efficiency gain over identity mapping.
Accelerating multijet-merged event generation with neural network matrix element surrogates
T. Herrmann et al. · SciPost Phys. 20 (2026) 071 · arXiv:2506.06203

Neural-network matrix-element surrogates accelerating multijet-merged event generation.
Differentiable Matrix Elements with MadJax
L. Heinrich & M. Kagan · J.Phys.Conf.Ser. 2438 (2023) 012137 · arXiv:2203.00057

MadJax: differentiable matrix elements, so generators can be trained through.
Reweighting and analysing event generator systematics by neural networks on high-level features
A. Furuichi, S. Lim & M. Nojiri · JHEP 07 (2025) 111 · arXiv:2503.01452

Reweighting and analysing generator systematics with neural networks on high-level features.
Mining gold from implicit models to improve likelihood-free inference
J. Brehmer et al. · Proc.Nat.Acad.Sci. 117 (2020) 5242 · arXiv:1805.12244

Mining gold: extracting joint likelihood ratios and scores from the simulator to power simulation-based inference.
MadMiner: Machine learning-based inference for particle physics
J. Brehmer et al. · Comput.Softw.Big Sci. 4 (2020) 3 · arXiv:1907.10621

MadMiner — the workhorse implementation of ML-based inference for particle physics.
Designing observables for measurements with deep learning
O. Long & B. Nachman · Eur.Phys.J.C 84 (2024) 776 · arXiv:2310.08717

Designing observables with deep learning (Long & Nachman): the observable itself becomes a learned object.
Observable optimization for precision theory: machine learning energy correlators
A. Bhattacharya, K. Fraser & M. Schwartz · JHEP 01 (2026) 151 · arXiv:2508.10988

Observable optimisation for precision theory: machine-learned energy correlators.
Attention Is All You Need
A. Vaswani et al. · arXiv:1706.03762

Attention is all you need — the transformer architecture behind current generative surrogates.

The Record · The Full Feed

Other sections: SMEFT & global fits · Multiboson & VBS · Boson polarization · Event generators · Precision calculations · Future colliders · Detectors & instrumentation.