Qcraft: A Modular Platform for Quantum Circuit Optimization and Surface Code Mapping via Reinforcement Learning

Abstract

Qcraft is a research-grade, modular desktop application for quantum circuit design, optimization, and surface code mapping. It leverages reinforcement learning (RL) to address the challenges of scalable, hardware-aware quantum compilation and error correction. This work presents the architecture, configurable workflows, and the novel RL-based surface code mapping module, highlighting the scientific motivation, reward function design, and future research directions.

---

1. Introduction

Quantum computing promises exponential speedups for certain problems, but practical realization is hindered by noise, limited connectivity, and hardware constraints. Surface codes are a leading error correction technique, but mapping logical qubits to physical hardware remains a complex, high-dimensional optimization problem. Qcraft addresses this by providing a unified, extensible platform for:

• Quantum circuit design and editing
• Hardware-aware circuit optimization
• RL-driven surface code mapping
• Artifact management and reproducibility
• Curriculum Learning: Progressive training with increasing difficulty, dynamic reward weighting, and robust convergence.
• Hardware Awareness: Supports IBM devices (IonQ in progress), native gate sets, and device-specific constraints.
• Modular and Configurable: YAML/JSON-driven configuration for all workflows, environments, and training parameters.
• Logging and Artifact Management: Automated tracking of training runs, metrics, and model artifacts for reproducibility.

---

Installation

Requirements

• Python: 3.9–3.11 (3.11 recommended)
• CUDA: 12.4 (required for RL training with surface code agents)
• Tested on: Linux, NVIDIA RTX 3070, CUDA 12.4, IBM Q devices

Install from PyPI

pip install qcraft

Installation

Option 1: Install from PyPI (Recommended)

pip install qcraft

Option 2: Install from GitHub Release Tarball

Download the latest qcraft-<version>.tar.gz from https://github.com/deba10106/Qcraft.git (see Releases tab), then install with:

pip install /path/to/qcraft-<version>.tar.gz

Note:

• Python 3.9–3.11 supported (3.11 recommended)
• CUDA 12.4 required for RL training

---

Usage

Main GUI

qcraft

Usage

To launch the Qcraft desktop application, simply run:

qcraft

---

Reward Functions: Overview

Surface Code Multi-Patch Agent

• Highly configurable reward function: Encourages valid mappings, hardware connectivity, adjacency, resource utilization, error minimization, and logical correctness.
• Curriculum learning: Dynamic reward weights and phase multipliers across training stages.
• See configs/multi_patch_rl_agent.yaml for all tunable parameters.

Circuit Optimization Module

• Reward engine: Penalizes gate count, depth, and SWAPs; rewards native gate usage and penalizes invalid gates.
• Curriculum learning: Difficulty and reward weights progress as training advances.
• See configs/optimizer_config.yaml for all tunable parameters.

---

Configuration and Customization

• All major workflows and RL environments are configured via YAML files in the configs/ directory.
• Surface Code Agent: configs/multi_patch_rl_agent.yaml
• Circuit Optimization Agent: configs/optimizer_config.yaml
• Device/Hardware: configs/ibm_devices.yaml, configs/ionq_devices.yaml
• Other: Logging, visualization, and more via their respective YAML files.

---

Packaging and PyPI Publishing

To build and publish your own version:

# Clean previous builds
rm -rf dist/*

# Build the package
python3 setup.py sdist bdist_wheel

# Check the package
pip install twine

# Upload to PyPI
twine upload dist/*

---

Support and Extensibility

• Qcraft is modular and extensible for new devices, reward functions, and optimization passes.
• Contributions and feedback are welcome for further research and development.

---

Citation

If you use Qcraft in academic work, please cite the corresponding paper or this repository.

---

For detailed technical documentation, architecture, and workflow explanations, please refer to the full README in the source repository.