PulserDiff: a pulse differentiable extension for Pulser

Vytautas Abramavičius, Melvin Mathé, Gergana Velikova, João P. Moutinho, Mario Dagrada, Vincent E. Elfving, Alexandre Dauphin, Joseph Vovrosh, Roland GuichardarXiv 2025 · DOI · arXiv

Abstract

Programming analog quantum processing units (QPUs), such as those produced by Pasqal, can be achieved using specialized low-level pulse libraries like Pulser. However, few currently offer the possibility to optimize pulse sequence parameters. In this paper, we introduce PulserDiff, a user-friendly and open-source Pulser extension designed to optimize pulse sequences over a well-defined set of control parameters that drive the quantum computation. We demonstrate its usefulness through several case studies involving analog configurations that emulate digital gates and state preparation. PulserDiff produces hardware-compatible pulses with remarkably high fidelities, showcasing its potential for advancing analog quantum computing applications.

Introduction

Neutral-atom quantum processing units (QPUs) primarily operate in an "analog" mode, where programmers control the many-body dynamics of atoms using laser pulses rather than discrete digital gates. Finding the optimal parameters for these pulses—amplitude, detuning, phase, and duration—to achieve a specific quantum state is a significant optimization challenge.

PulserDiff addresses this by integrating the popular Pulser library with the PyTorch ecosystem. By making the entire pulse-building pipeline differentiable, it enables researchers to use automatic differentiation (AD) to optimize laser pulse sequences and atom layouts seamlessly. This transition from manual trial-and-error to automated optimization allows for the creation of hardware-ready, high-fidelity quantum operations with minimal tuning.

Methods: Differentiable Pulse Programming

The core innovation of PulserDiff is its integration with PyTorch. It wraps standard Pulser objects into a new QuantumModel class, which subclasses torch.nn.Module. This allows standard machine learning optimizers, such as Adam, to find the best configurations automatically.

Key methodological features include:

  • Differentiable Parameters: Pulse amplitudes ($\Omega$), detunings ($\delta$), phases ($\phi$), and even inter-atomic distances ($r_{ij}$) are defined as PyTorch tensors with gradients enabled.
  • Hardware-Aware Constraints: The library enforces physical bounds on parameters, ensuring optimized pulses remain executable on real hardware.
  • Smooth Waveform Interpolation: To move beyond idealized square pulses, PulserDiff implements sine-interpolation. This creates continuous waveforms that are more hardware-amenable while remaining fully differentiable.
Example of a differentiable pulse sequence showing Rabi frequency ($\Omega$) and detuning ($\delta$) profiles.

System Design and Benchmarks

One of the most powerful ideas in PulserDiff is treating the physical layout of the atoms as an optimizable parameter. This allows for the co-design of hardware (the register geometry) and software (the pulses) for specific tasks. The tool was evaluated using simulated benchmarks on registers ranging from 2 to 7 qubits in various geometries, including linear, rectangular, and triangular layouts.

A two-atom register layout ($q0, q1$) used for gate emulation benchmarks.

Key Results

The optimization capabilities of PulserDiff yield significant performance gains over traditional methods:

  • High-Fidelity Gate Emulation: For a 2-qubit system, the tool achieved 99.99% fidelity in emulating a global Hadamard gate.
  • Superior Scaling: In a 7-qubit system, optimized custom continuous pulses maintained fidelities above 99.5%, significantly outperforming simple square pulses (~95%).
  • Robust State Preparation: In a 6-qubit linear register, PulserDiff achieved 99.85% fidelity in target state preparation, effectively mitigating unwanted long-range interactions that normally disrupt dynamics.

Conclusion and Future Work

PulserDiff provides a clear path for running digital-style algorithms on current-generation analog hardware. While currently limited by the computational expense of full state-vector simulation, future versions aim to integrate with high-performance emulators based on Matrix Product States (MPS) to support dozens or hundreds of qubits. Additionally, future work will focus on discovering pulses that are inherently resilient to hardware noise and decoherence.

Summary updated

Abstract

Programming analog quantum processing units (QPUs), such as those produced by Pasqal, can be achieved using specialized low-level pulse libraries like Pulser. However, few currently offer the possibility to optimize pulse sequence parameters. In this paper, we introduce PulserDiff, a user-friendly and open-source Pulser extension designed to optimize pulse sequences over a well-defined set of control parameters that drive the quantum computation. We demonstrate its usefulness through several case studies involving analog configurations that emulate digital gates and state preparation. PulserDiff produces hardware-compatible pulses with remarkably high fidelities, showcasing its potential for advancing analog quantum computing applications.
BibTeX
@misc{abramavicius2025pulserdiff,
  author = {Vytautas Abramavičius and Melvin Mathé and Gergana Velikova and João P. Moutinho and Mario Dagrada and Vincent E. Elfving and Alexandre Dauphin and Joseph Vovrosh and Roland Guichard},
  title = {{PulserDiff: a pulse differentiable extension for Pulser}},
  year = {2025},
  eprint = {2505.16744},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.16744},
  howpublished = {arXiv},
}
Figure