⚛️ quantum physics

Inverse Design of Quantum Control Sequences with Fourier Neural Operators

This paper introduces a Fourier Neural Operator-based framework that accelerates the inverse design of quantum control sequences by over six orders of magnitude, enabling the efficient purification of high-dimensional molecular states where conventional optimization methods are computationally intractable.

Original authors: Anastasia Pipi, Valentin Duruisseaux, Emily Been, Xuecheng Tao, Taylor L. Patti, Anima Anandkumar, Prineha Narang

Published 2026-08-05

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Anastasia Pipi, Valentin Duruisseaux, Emily Been, Xuecheng Tao, Taylor L. Patti, Anima Anandkumar, Prineha Narang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to steer a tiny, invisible boat through a stormy ocean. In the world of quantum physics, this "boat" is a molecule, and the "storm" is the chaotic dance of its internal energy levels. Scientists want to guide these molecules into a perfectly calm, specific state so they can be used for ultra-precise measurements or as building blocks for future quantum computers. This process is called "quantum control." The problem is that molecules are incredibly complex; they have so many possible ways to wiggle and vibrate that the number of possibilities grows so fast it becomes impossible to calculate the perfect steering path using traditional math. It's like trying to find the single best route through a maze that has more paths than there are grains of sand on Earth. If you try to simulate every possible turn on a computer, it takes so long that you might as well give up. This is the bottleneck that has held back progress in making quantum technologies work with complex molecules.

Now, imagine you have a magical, super-fast crystal ball that can predict exactly how the boat will move through the storm without you having to do the heavy math yourself. That is essentially what this paper introduces. The researchers, working with a molecule called hydronium (which is just water with an extra hydrogen ion, $H_3O^+$ ), have built a new type of artificial intelligence called a "Fourier Neural Operator" (FNO). Think of this FNO as a master chef who has tasted every possible dish (every possible laser pulse) and learned the secret recipe for how the ingredients (the molecule's energy states) will change. Instead of cooking every single dish from scratch to see how it turns out—which takes hours—the FNO predicts the result in a flash. Using this AI, the team created a "stochastic pulse-measurement planner" (a fancy name for a smart game plan) that designs a sequence of laser pulses to clean up a messy, hot mixture of molecular states and turn it into a single, pure state. They tested this on a simulation of the hydronium molecule, which has a massive "playground" of 888 different energy states. The result? The AI found a way to purify the molecule with a 98% success rate, doing the work 10 million times faster than the old, slow computer methods. It's like going from taking ten hours to solve a puzzle to doing it in ten minutes, all while finding a better solution than the previous best attempts.

The Story of the Quantum Boat and the Magic Crystal Ball

The Setup: A Messy Molecule
Picture a molecule of hydronium ( $H_3O^+$ ) trapped in a cage of light. Inside, it's not sitting still; it's buzzing with energy, bouncing around in a "mixed" state. It's like a bag of marbles where some are red, some are blue, and some are green, all jumbled together. To use this molecule for science, we need to sort the marbles until they are all one color—a "pure" state. To do this, scientists use lasers (the "steering wheel") to push the molecule from one energy level to another. But because the molecule has so many levels (888 of them in this study), figuring out the exact sequence of laser pushes is a nightmare. If you try to calculate the path for every possible laser setting, your computer would need to run for days or even years.

The Old Way: The Slow Cook
Before this paper, scientists used a method called "direct numerical propagation." Imagine trying to predict the weather by manually calculating the movement of every single air molecule in the atmosphere. It's accurate, but it's painfully slow. In the quantum world, this meant simulating the molecule's reaction to a laser pulse over and over again. For a single sequence of pulses, this could take about 10 hours. It was so slow that trying to find the best sequence was practically impossible.

The New Way: The FNO Crystal Ball
The authors introduced a new tool: the Fourier Neural Operator (FNO). Instead of calculating every tiny step of the molecule's movement, the FNO learns the rules of how the molecule moves. It's like a student who watches a thousand physics problems and then learns the underlying pattern, allowing them to solve a new problem instantly without doing the long math.

The researchers trained this FNO on the hydronium molecule. They fed it data about how the molecule reacts to different laser frequencies and polarizations (the direction the laser waves wiggle). Once trained, the FNO could predict the molecule's future state in a single "forward pass."

The Speed: The paper reports that the FNO is up to $1.84 \times 10^7$ (18.4 million) times faster than the old method when looking at large batches of possibilities.
The Accuracy: The predictions were incredibly close to the real math, with an error rate (infidelity) of less than $3 \times 10^{-3}$ . This means the crystal ball was right almost every time.

The Game Plan: FNO-SPMP
Having a fast crystal ball is great, but you still need a strategy to win the game. The team built a "stochastic pulse-measurement planner" (FNO-SPMP). Here is how it plays the game:

Sampling: It looks at many possible laser pulses (different frequencies and durations) and asks the FNO: "If we hit the molecule with this pulse, what happens?"
Scoring: It rates the pulses based on how well they move the molecule toward the "pure" state.
The Twist (Measurement): In the quantum world, you can't just push the molecule; you have to "check" it. After a pulse, the team simulates a measurement. If the molecule is in the right "branch" of possibilities, great! If not, the plan adjusts. The FNO-SPMP is smart enough to handle these random outcomes, building a tree of possibilities without getting lost.
Refinement: Once it picks a sequence, it uses the FNO's ability to do "gradient descent" (a mathematical way of fine-tuning) to tweak the laser settings slightly for perfection.

The Results: A Clean Sweep
The team tested this on a simulation of the hydronium molecule at a temperature of 20 K (very cold, but not absolute zero).

Success Rate: The FNO-SPMP successfully purified the molecule to a target state with 0.98 (98%) population in the desired state.
Reliability: It achieved this in 86.2% of the attempts (sequence-success rate).
Efficiency: It used about 26 to 27 pulses on average to reach the goal.
Speed of Design: While the old Reinforcement Learning (RL) methods took about 10 hours to find a decent plan, the FNO-SPMP did it in 10 to 20 minutes. That is a 30 to 60 times improvement in finding the solution.

What This Means
The paper explicitly shows that for complex molecules with huge "Hilbert spaces" (the mathematical space of all possible states), the old way of brute-force calculation is too slow. The FNO approach doesn't just speed things up; it makes the impossible possible. It allows scientists to design control sequences for systems that were previously too big to handle. The authors note that while they used simulations (using a powerful computer to mimic the real world), the results suggest that this method could be a game-changer for real-world experiments, potentially helping to build better quantum sensors and computers. They didn't just find a slightly faster way to do the same thing; they found a way to do a task that was effectively out of reach before.

In short, by teaching a computer to "feel" the quantum rules rather than just crunching numbers, the researchers have opened the door to steering the most complex molecules with the precision of a surgeon, all in the blink of an eye.

Technical Summary: Inverse Design of Quantum Control Sequences with Fourier Neural Operators

Problem Statement
Quantum optimal control is essential for steering quantum dynamics, particularly for preparing specific states in complex systems like polyatomic molecular ions. However, the computational cost of designing control protocols scales rapidly with the dimension of the Hilbert space. For systems such as the hydronium ion ( $H_3O^+$ ), the internal structure creates a high-dimensional state space where the number of possible control sequences and states grows exponentially. Traditional gradient-based methods (e.g., GRAPE, Krotov) and parameterized-pulse algorithms (e.g., CRAB) require repeated numerical propagation of the quantum state, becoming computationally prohibitive for large molecular systems. Furthermore, existing machine learning approaches, such as reinforcement learning (RL), often struggle with scalability, require extensive training times, and are frequently limited to restricted discrete action spaces, hindering the exploration of continuous control parameters necessary for high-precision spectroscopy.

Methodology
The authors propose a framework centered on a Fourier Neural Operator (FNO) to serve as a fast, differentiable surrogate for quantum dynamics, enabling the inverse design of control protocols.

System and Hamiltonian: The study focuses on Quantum Logic Spectroscopy (QLS) of the hydronium ion ( $H_3O^+$ ) co-trapped with a calcium ion ( $Ca^+$ ). The system is modeled in an 888-dimensional subspace (444 molecular states $\times$ 2 motional manifolds). The control involves two-photon Raman sideband pulses driven by electric fields with specific frequencies ( $\omega$ ) and polarizations ( $\sigma \in \{\sigma_+, \sigma_-\}$ ).
FNO Surrogate Construction:
- Architecture: The FNO learns mappings between functions in Fourier space, allowing it to predict time-dependent dynamics in a single forward pass.
- Physics-Informed Embedding: Instead of treating control parameters as static scalars, the input is embedded as time-dependent oscillatory channels based on the detuning of the laser frequency from molecular transitions ( $\Delta_{J'J}$ ). This captures the phase accumulation and resonance conditions directly.
- Symmetry Utilization: The model is trained on the $\sigma_+$ polarization channel. The $\sigma_-$ channel dynamics are inferred using a frequency-reflection relation and state-index permutation, avoiding the need for separate training.
- Loss Function: An activity-weighted loss function is employed, prioritizing errors in basis-state populations that undergo significant dynamical changes, ensuring high fidelity for active transitions.
Inverse Design Protocol (FNO-SPMP):
- The authors introduce the FNO-guided Stochastic Pulse-Measurement Planner (FNO-SPMP). This protocol aims to purify an initially mixed Boltzmann distribution into a single target state.
- Process: At each control step, the planner scores candidate pulses (frequency, duration, polarization) based on selective transfer, branch purity, and excited-branch population probability.
- Stochastic Planning: A pool of top-scoring candidates is formed, and a pulse is sampled stochastically to build a sequence. The planner tracks the $\nu=0$ motional branch as the primary trajectory while accounting for $\nu=1$ outcomes.
- Gradient Refinement: Due to the differentiability of the FNO, continuous control parameters ( $\omega, \tau$ ) of selected pulses can be locally refined via gradient descent to optimize the inverse-design loss.
- Validation: Generated sequences are validated via Monte Carlo rollouts using direct numerical propagation (CUDA-Q) to ensure independence from surrogate prediction errors.

Key Contributions

First FNO-based Surrogate for Quantum Control: The paper introduces the first FNO-based surrogate model specifically designed for the inverse design of quantum control sequences.
Conditional Dynamics Inference: The FNO successfully infers quantum dynamics conditioned on experimental control parameters, predicting full population trajectories for candidate Raman sideband pulses in an 888-dimensional subspace.
FNO-SPMP Protocol: The development of a stochastic planner that leverages the FNO's speed and differentiability to score, select, and refine control pulses, enabling exploration of a larger, continuous control space.
Demonstration on Hydronium: The framework is demonstrated on the $H_3O^+$ molecule, achieving high-fidelity state preparation in a regime previously inaccessible to direct optimization.

Results

Accuracy: The FNO surrogate achieves high precision, with average population infidelities consistently below $3 \times 10^{-3}$ across pulse times, drive frequencies, and random mixed initial states.
Computational Speedup: The FNO provides a massive acceleration over GPU-accelerated numerical propagation (CUDA-Q). For batched forward evaluations, speedups range from $5.13 \times 10^4$ to $1.84 \times 10^7$ depending on batch size and evaluation type (state-batch vs. frequency-batch).
Inverse Design Performance:
- In an 888-dimensional subspace at 20 K, the FNO-SPMP achieves a target-state population of 0.98 with a sequence-success rate of up to 86.2%.
- Comparison with RL: In a shared discrete control space, FNO-SPMP achieves nearly twice the success rate (79.4% vs. 42.8%) of a reinforcement learning baseline (RL-QLS) while using roughly half as many pulses (27.9 vs. 49.0).
- Efficiency: The time to generate a high-performing protocol is reduced from approximately 10 hours (for RL training and validation) to 10–20 minutes for FNO-SPMP, representing a 30–60× reduction in generation time.

Significance and Claims
The paper claims that operator-learning surrogates, specifically FNOs, can enable inverse design in quantum systems with Hilbert spaces too large for conventional direct optimization. By replacing costly numerical propagation with a differentiable surrogate, the framework allows for the efficient exploration of continuous control spaces and the optimization of complex pulse sequences. The authors position this work as a step toward combining operator-learning frameworks with optimal quantum control in high-dimensional systems, suggesting applicability to quantum computing and sensing where robust control and exact simulation are challenging. The work does not claim to have solved all quantum control problems but demonstrates a viable path for scaling control design to molecular systems where traditional methods fail due to computational cost.

The Story of the Quantum Boat and the Magic Crystal Ball

More like this