preparation. (b)Propagating single-pulse molecular dynamics with the FNO.
A multi-institution research collaboration led by UCLA (NarangLab), Caltech, and NVIDIA has introduced a machine-learning framework to automate the inverse design of quantum control pulse sequences. Detailed in a preprint published on arXiv (arXiv:2608.03702) and presented at IEEE Quantum Week 2026, the method utilizes a Fourier Neural Operator (FNO) to learn high-dimensional molecular quantum dynamics, replacing numerical differential equation solvers inside optimal control loops.
The FNO surrogate was trained on GPU-accelerated quantum state propagations generated by NVIDIA CUDA-Q Dynamics across an 888-dimensional Hilbert space modeling the trapped hydronium ion (H₃O⁺). Incorporating physics-informed frequency detuning embeddings and polarization symmetry constraints (σ⁺ and σ⁻ channels), the FNO surrogate predicts population trajectories across Raman sideband pulse windows up to ~10⁷× faster than accelerated GPU numerical propagation solvers. The surrogate is fully differentiable, enabling direct gradient-based optimization of continuous laser pulse parameters.
| [ FNO-SPMP Control Performance vs. Reinforcement Learning Baseline ] | ||
|---|---|---|
| Control Metric | Fourier Neural Operator (FNO-SPMP) | Standard Reinforcement Learning (RL) |
| Target State Fidelity & Success | • Target Population Density: 0.98 • Preparation Success Rate: Up to 86.2% | • ~43% Success Rate Baseline • Higher Residual Thermal Entropy |
| Sequence Overhead | • ~50% Reduction in Required Pulse Count | • 2× Larger Optical Pulse Overhead |
| Synthesis Compute Latency | • 10 to 20 Minutes (Differentiable FNO) | • ~10 Hours (Discrete RL Search) |
Built on top of the differentiable FNO, the team implemented a stochastic pulse-measurement planner (FNO-SPMP) to navigate thermal population distributions at 20 K. On H₃O⁺, the framework generated pulse sequences that funneled thermal populations into single rotational/hyperfine states with an 86.2% success rate, reducing sequence synthesis latency from 10 hours down to 10–20 minutes. The methodology establishes an operator-learning paradigm within NVIDIA’s CUDA-Q platform to synthesize pulse-level control for molecular spectroscopy and complex QPU gate calibration.
Review the technical update via UCLA NarangLab here, examine the peer-reviewed preprint on arXiv here, and read our prior coverage of NVIDIA CUDA-Q Platform Integrations here.
September 14, 2026
