Rajkumar Buyya's QLSTM Beats Classical AI on Particle Physics With 300 Parameters

A hybrid quantum-classical LSTM with fewer than 300 parameters gets within about 1% of deep-learning AUC on a particle-physics benchmark while training on only 20K rows.

·
·
·
Rajkumar Buyya's QLSTM Beats Classical AI on Particle Physics With 300 ParametersPRO
  • A new paper combines quantum circuits with an LSTM (QLSTM) and trains it with federated learning on SUSY physics data.
  • QLSTM reaches 0.880 AUC with fewer than 300 parameters and 20K training rows.
  • That is within about 1% of deep-learning baselines trained on 5M rows with roughly 300K parameters.
  • It beats an in-paper classical LSTM (0.862) and a plain VQC (0.823) on the same data.
  • AUC stays almost flat as nodes are added, and dropping from 18 to 7 features costs only about 0.5%.
  • Limits: everything ran in a simulator, data was split evenly (IID) from one source, and no code was released.

Researchers working with Rajkumar Buyya’s group have combined quantum circuits, long short-term memory networks and federated learning for particle-physics classification. In the preprint, they train copies of a quantum LSTM on separate data partitions and aggregate the local models into one global classifier.

On 20,000 events sampled from the five-million-row SUSY dataset, the strongest QLSTM reaches an area under the curve of 0.880 with fewer than 300 trainable parameters. A three-node federated version scores 0.878. Both results overlap the 0.87 to 0.89 range reported by much larger neural networks in the original SUSY benchmark, although that cross-paper comparison does not control for training conditions or computational cost. The manuscript states that it has been accepted at IEEE WCCI 2026.

Collider data meets a tiny model

The SUSY dataset is a standard high-energy physics benchmark containing five million simulated collision events. Its binary classification task separates events associated with supersymmetry from background events. Supersymmetry extends the Standard Model with a proposed relationship between fermions and bosons.

Each event contains 18 measured and derived kinematic features. The paper evaluates models with all 18 features and with a reduced set of seven, testing whether the classifiers retain useful performance after most inputs are removed.

Quantum execution constrains the experiment. Current noisy intermediate-scale quantum devices have few reliable qubits, while classical simulation becomes increasingly expensive as qubit counts grow. Repeatedly evaluating parameterized circuits over millions of events would exceed the practical scope of the six-qubit simulation used here.

Federated learning distributes training among multiple nodes. Each node processes a local data partition and sends model updates for aggregation, a design that can support data locality across laboratories or detectors. The paper simulates that arrangement by dividing one source dataset into statistically similar partitions; it does not operate across physical sites or quantum processors.

Six qubits inside each LSTM gate

An LSTM is a recurrent neural network whose input, forget and output gates control how information enters, persists in and leaves an internal memory state. The QLSTM replaces the usual trainable transformations within those gates with variational quantum circuits, or VQCs.

A VQC contains quantum gates with adjustable rotation angles. A classical optimizer updates those angles during training in much the same way that it updates neural-network weights. Each circuit in this study uses six qubits, angle-based input encoding and four trainable layers. Entangling operations allow the qubits to represent correlations among encoded values.

Pro article

This story is for Pro members

You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.

Trending
  • No trending articles

Comments

avatar

Next Reads