QINNs: Quantum-Informed Neural Networks
Aritra Bal ∗, Markus Klute, Benedikt Maier ∗, Melik Oughton, Eric Pezone, Michael Spannowsky ∗ · 2025
Read QINNs: Quantum-Informed Neural Networks by Aritra Bal ∗, Markus Klute, Benedikt Maier ∗, Melik Oughton, Eric Pezone, Michael Spannowsky ∗ (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-17984.txt.
Introduction
Aritra Bal and five co-authors bring ideas from quantum information into plain classical neural networks for particle physics. Each particle in a collision is treated like a qubit, and a quantum Fisher information matrix sums up how the particles relate. That compact summary then feeds a graph network that sorts jets of particles.
The quantum-based features show distinct patterns for plain jets and for top quark jets, matching physical hopes. Sorting improved as a result. The authors frame the method as a broad, readable, and scalable route to quantum-informed study of collisions, without needing an actual quantum computer to run it.
Catalog record
| CREATOR | Aritra Bal ∗, Markus Klute, Benedikt Maier ∗, Melik Oughton, Eric Pezone, Michael Spannowsky ∗ |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-17984
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
QINNs: Quantum-Informed Neural Networks Aritra Bal ∗, Markus Klute, Benedikt Maier ∗, Melik Oughton, Eric Pezone, Michael Spannowsky ∗ Abstract Classical deep neural networks can learn rich multi-particle correlations in collider data, but their inductive biases are rarely anchored in physics structure. We propose quantum-informed neural networks (QINNs), a general framework that brings quantum information concepts and quantum observables into purely classical models. While the framework is broad, in this paper, we study one concrete realisation that encodes each particle as a qubit and uses the Quantum Fisher Information Matrix (QFIM) as a compact, basis-independent summary of particle correlations. Using jet tagging as a case study, QFIMs act as lightweight embeddings in graph neural networks, increasing model expressivity and plasticity. The QFIM reveals distinct patterns for QCD and hadronic top jets that align with physical expectations. Thus, QINNs offer a practical, interpretable, and scalable route to quantum-informed analyses, that is, tomography, of particle collisions, particularly by enhancing well-established deep learning approaches. \preprintA IPPP/25/60 1 Introduction Jets, collimated cascades of hadrons originating from energetic partons, provide one of the most powerful microscopes with which the LHC explores quantum chromodynamics (QCD) and searches for new physics Marzani:2019hun . Their internal substructure encodes a wealth of information about the underlying hard process, colour flow, and hadronisation dynamics Gallicchio:2010sw ; Larkoski:2013eya ; Larkoski:2014gra . When the initiating parton is a top quark with TeV-scale transverse momentum, its three hadronic decay products overlap so tightly that they are reconstructed as a single top jet Butterworth:2008iy ; Kaplan:2008ie ; Plehn:2009rk . Disentangling the physics encoded in such highly correlated multi-particle radiation patterns is central to boosted-object tagging and precision top-quark measurement Soper:2012pb ; Soper:2014rya ; Larkoski:2017jix ; Kogler:2018hem . Yet, the complexity of these correlation structures resists simple factorized descriptions and often leaves modern machine-learning classifiers difficult to interpret physically Cogan:2014oua ; Komiske:2018cqr ; Qu:2019gqs ; Shlomi:2020gdn . Jet tomography seeks to overcome this challenge by treating the energy flow inside a jet as a tomographic object, much like medical imaging reconstructs structure from multiple projections Dreyer:2018tjj . We introduce quantum-informed neural networks (QINNs), a general framework that brings quantum information concepts and observables into purely classical models; in this paper, we instantiate it as quantum-informed jet tomography built around the QFIM. The QFIM provides a basis-independent measure of the dynamical sensitivity of a quantum state and thus naturally encodes correlations between jet constituents when events are embedded in a resource-efficient one-particle-one-qubit (1P1Q) representation Braunstein:1994zz ; Liu:2019xzz . Importantly, we find that the QFIM extracted from ensembles of QCD and top jets exhibit strikingly different structures: QCD jets appear largely diagonal, reflecting quasi-independent splittings, whereas top jets show pronounced off-diagonal elements characteristic of their three-pronged colour-connected topology. This makes the QFIM an interpretable observable for jet physics. More broadly, we use QINN as an umbrella term for methods that inject quantum structure into classical learning while remaining fully classical at training and inference. A QINN is specified by an encoding of event information into a quantum state, for example, 1P1Q or continuous-variable maps, a choice of quantum observables or metrics, for example, the QFIM, reduced density matrices, entanglement measures, or quantum distances, and an interface to classical models, for example inputs, edge or node embeddings, regularisers, or initialization priors. This work studies one concrete member of that family: a 1P1Q encoding with the QFIM used both as an interpretable observable and as guidance for optimisation. Recent progress in jet tagging has largely resulted from deep learning methods that learn correlations directly from data. However, these models often remain opaque and can be sensitive to training choices. Alternative approaches, including energy-flow polynomials, correlation functions, and graph-based neural networks, aim to make the underlying structures more transparent. In parallel, the growing field of quantum machine learning has introduced representations such as amplitude and qubit encodings to capture relations in high-dimensional data Havlicek:2018nbf ; Schuld:2018hrv . Our study positions itself at the intersection of these developments: by using QFIM as an interpretable, quantum-native probe, we connect quantum geometry with classical machine learning to expose and exploit the correlation patterns that distinguish different jet topologies. In this context, QFIM also links naturally to optimisation via its role as a quantum metric and in natural-gradient training, and connects to known trainability phenomena in variational circuits Stokes:2019ads ; McClean:2018bwh ; Preskill:2018 ; Blance:2020nhl . Thus, we first establish the role of QFIM as an untrained classifier, then demonstrate how it can inform and accelerate a simple neural network, and finally extend the framework to GNNs enriched with both QFIM-derived and kinematic features. This paper makes three contributions. First, it formulates QINN as a general framework for quantum-informed design of classical models. Second, it develops a QFIM-based jet tomography instantiation that exposes class-dependent correlation geometry in an interpretable way. Third, it shows how QFIM-derived quantities act as lightweight graph embeddings and training priors that improve stability and performance in GNNs. Concretely, in Sec 2 we introduce the QFIM and review the kinematic anatomy of boosted hadronic top decays, set out the qubit map and detail the construction of the QFIM as a correlation matrix. In Sec. 3 , we show that the QFIM can serve directly as a classifier without training. By vectorising the averaged QFIM over event samples, we obtain class prototypes for QCD and top jets. Each new event is then scored according to its proximity to these prototypes, yielding a powerful discriminant even in its untrained form. Building on this, we convert the QFIM-derived feature vector into the input of a shallow neural network with a single trainable encoding layer and linear output. This model learns patterns that closely track the QFIM structure itself, and crucially, initialising the network with the QFIM pattern accelerates convergence and improves optimisation dynamics. In this way, the QFIM does not just separate classes, but also informs the training of neural networks. Having established the QFIM as a tomographic probe and an informative prior for machine learning, we extend the framework to graphs and GNNs in Sec. 4 . In this representation, jet constituents are nodes and the QFIM entries act as edge features, complemented by kinematic information stored on the nodes. As shown in Sec. 5 , the resulting quantum-informed GNN captures both the geometry of quantum correlations and the kinematics of the jet, achieving excellent tagging performance and exhibiting faster and smoother training than purely kinematics-based models Qu:2019gqs ; Shlomi:2020gdn . Thus, the QFIM acts both as a physically interpretable observable and as a guiding principle for network design. Finally, we summarise our conclusions in Sec . 6 . Our study demonstrates that QINNs for jet tomography unify the interpretability of quantum geometry with the scalability of modern deep learning. By embedding jets into the 1P1Q formalism and exploiting the QFIM, we expose correlation structures that distinguish top jets from the QCD background, accelerate training through informed initialisation, and enhance GNN performance. This study outlines how to integrate quantum observables into collider analyses for classification tasks. It is compatible with near-term quantum hardware and outlines an approach for how to utilise hybrid classical–quantum analysis frameworks. 2 Subjet-correlations of boosted hadronic jets We want to quantify, in a basis–independent manner, how much of the