00Quantum machine-learning research

QMANN

QMANN combines classical neural networks with quantum memory operations to push learning capacity past what classical models alone can reach. The reference implementation provides practical quantum-inspired algorithms suitable for near-term NISQ devices, with three reproducible modes — Theoretical, Simulation, and Hardware.

View the repository ↗QMANN v2.0 preview ↗
Execution modesCost gate before hardware
01

Theoretical

Unlimited (ideal quantum computer)

Free
02

Simulation

Up to 20 (with noise modelling)

Free
03

Hardware

4–12 (NISQ-bounded)

Paid

Hardware mode runs the cost estimator before shots are submitted.

Reproducible modes
3
Qubits in simulation
20
Qubits on hardware
4–12
Open licence
Apache-2.0

01Execution modes

Most QML libraries hide the cost surface and the noise model. QMANN exposes both, and ships three explicit modes so you always know what your run actually means.

01

Free

Validate algorithm structure and theoretical bounds before any cost is incurred. Perfect gates, infinite coherence.

Qubits
Unlimited (ideal quantum computer)
Audience
Research papers, theoretical bounds, algorithm design
02

Free (compute only)

Classical simulation with noise modelling and quantum-inspired operations. Ship the science before paying for shots.

Qubits
Up to 20 (with noise modelling)
Audience
Algorithm validation, development, education, reproducible research
03

Paid (IBM ~$0.001 / shot, IonQ ~$0.01 / shot)

Run on real quantum devices with cost-estimator gating. Hardware mode requires explicit credentials and a confirmed budget.

Qubits
4–12 (NISQ-bounded)
Audience
Proof-of-concept, hardware benchmarking, quantum advantage validation

QMANN now sits inside a clearer professional context: theoretical mode, simulation mode, hardware budgets, and repeatable evidence for quantum-machine-learning programs.

3 modes
20-qubit simulation
Hardware budget gate
Researchers in laboratory coats discussing findings in a high-tech research lab

FIG 02 · Quantum ML operations — from theoretical bounds to governed hardware runs

Simulation-to-hardware research pathA hardware run is an evidence decision, not the default execution mode.
  1. 01input

    Define the learning question, classical baseline, target metric, and acceptable compute budget.

    Input · question + baseline

  2. 02process

    Use theoretical mode to test algorithm structure and expected limits without claiming hardware behaviour.

    Artifact · analytical bounds

  3. 03process

    Run the controlled noise model and compare the result with the classical baseline before spending shots.

    Evidence · reproducible run

  4. 04gate

    Review qubit count, provider route, credentials, estimated shot cost, and the reason hardware is necessary.

    Gate · explicit budget

  5. 05output

    Compile accepted hardware work through qmesh and retain the signed run context, result, and limitations.

    Output · provenance record

01

Quantum memory layer with entangled qubit registers.

02

Hybrid controller with parameterised quantum circuits.

03

Advanced error mitigation and measurement strategies.

04

Healthcare, industrial IoT, autonomous systems modules.

Every technique below maps to a published reference. Pull the paper, read the section, then run the code.

01

QSegRNN-style segment-based quantum recurrent neural networks. Quantum gates for LSTM operations, hybrid memory cells, enhanced temporal pattern recognition.

EPJ Quantum Technology · March 2025

02

Conditional diffusion-based parameter generation. Classical solution warm-start, adaptive bias correction. Faster convergence with reduced optimisation overhead.

Physical Review · 2025; EPJ Quantum Technology · August 2025

03

Grover-inspired amplitude amplification for optimisation problems. O(√N) speedup for unstructured search with adaptive oracle construction.

Cornell Lawler Research · January 2025

04

Quantum-classical attention integration. Multi-head quantum attention, entanglement-based correlations, configurable quantum attention ratio (0–100%).

arXiv:2504.00068 · arXiv:2501.15630

05

Enhanced virtual distillation with noise resilience. Multiple virtual copies, adaptive thresholds, maintains effectiveness with imperfect gates.

Communications Physics · October 2024

06

ML models for quantum error prediction and correction. Neural network error models, adaptive strategies, real-time error prediction and correction.

Recent QEM literature · 2025

A1Application and stack annexSector pipelines and qmesh provenance path

Sector — example pipelines

Finance
QAOA portfolio optimisation, fraud detection on transaction streams, market prediction with quantum LSTMs, quantum-enhanced risk modelling and stress testing.
Drug discovery
Quantum transformers for molecular property prediction, drug-target binding affinity, Grover-dynamics molecule generation, ADMET safety prediction.
Materials science
Quantum property prediction, adaptive-VQE crystal structure refinement, battery material design, catalyst design via quantum simulation.
Healthcare
Pattern recognition for diagnosis, optimisation for personalised medicine, protein folding via quantum simulation, clinical-trial patient matching.
Industrial
Quantum anomaly detection for predictive maintenance, classifiers for defect detection, QAOA for logistics, quantum simulation for chemical processes.

When you target hardware, QMANN circuits flow through qmesh for IR lowering, error mitigation, and ed25519-signed run manifests. The same provenance chain that satisfies regulated industries also gives ML researchers fully reproducible benchmarks.

06Quick start

qmann / examplesSH
$ git clone https://github.com/neuraparse/QMANN.git
$ cd QMANN
$ pip install -e .

# 🔬  Theoretical mode (FREE)
$ python examples/01_theoretical_mode.py

# 💻  Simulation mode (FREE)
$ python examples/02_simulation_mode.py

# ⚛️  Hardware mode (PAID — estimate first!)
$ python scripts/estimate_hardware_costs.py --qubits 6 --shots 1000
$ python examples/03_hardware_mode.py

Open research

GitHub ↗Compile through qmesh →