Researchers at UCLA and Caltech, working with NVIDIA, have developed a new approach to controlling molecules using artificial intelligence. The team trained a Fourier Neural Operator to learn the quantum dynamics of hydronium (H₃O⁺), a molecule whose subtle behavior may hold clues to physics beyond the Standard Model.
This allows the design of control protocols for molecules in state spaces previously too complex for conventional optimization methods. “We need a model that has actually learned the underlying physics,” explains the UCLA NarangLab group, highlighting the limitations of large language models in the quantum realm, and announcing the work at IEEE Quantum Week as part of NVIDIA’s platform launch.
Fourier Neural Operator Learns Hydronium Molecular Dynamics
Conventional methods for controlling molecules falter when faced with expansive state spaces too large for effective optimization, prompting the development of a learned surrogate model for quantum dynamics. This new technique bypasses the need for repeated, computationally expensive numerical simulations within the optimization loop by training a Fourier Neural Operator to predict molecular behavior.
The core of this work, detailed in “Inverse Design of Quantum Control Sequences with Fourier Neural Operators,” involves training the Fourier Neural Operator to act as a simulator, accepting the molecule’s current population distribution and laser pulse parameters to predict the resulting population trajectory in a single pass. A key design choice was a physics-informed embedding, where laser frequencies are encoded as detunings from relevant molecular transitions, enhancing the model’s efficiency and accuracy.
Because experiments measure a shared motional mode, the planner must independently track molecular populations, a task made tractable by the FNO’s predictive capabilities. On hydronium, the FNO-SPMP approach, a combination of the Fourier Neural Operator and a sequential Monte Carlo planning method, achieves a target-state population of 0.98 with a success rate of up to 86.2%. This performance represents an improvement over traditional optimal-control methods, which struggle with the dimensionality of the problem.
The team’s success is particularly notable given the sensitivity of hydronium’s inversion transitions to variations in fundamental constants, making it a promising candidate for detecting subtle deviations from established physical laws. NVIDIA’s infrastructure played a critical role in enabling this research, providing the computational power necessary to train and deploy the Fourier Neural Operator.
The company’s CUDA Quantum platform and DGX Quantum systems offer both quantum computing services and the GPU-accelerated computing platform essential for hybrid quantum-classical workflows. This collaboration builds on NVIDIA’s recent advancements in quantum calibration models, reported on July 28, 2026 to function across six qubit modalities, and the January 6, 2026 launch of NVQLink, an open architecture for low-latency integration of quantum processors with GPU supercomputers.
NVIDIA’s QEC-powered CUDA-Q Logical compiler facilitates error-corrected quantum chip programming, streamlining the development process. The researchers emphasize that simply prompting a system for a control pulse sequence is insufficient; a truly effective solution requires a model capable of understanding and predicting molecular dynamics. This work demonstrates the potential of machine learning to address complex challenges in quantum control.
Physics-Informed Embedding Accelerates Quantum Control Simulation
Researchers report that this method significantly improves the speed and success rate of designing control protocols for molecules with numerous rotational and hyperfine levels. Conventional optimization methods struggle with polyatomic hydronium because preparing it for precision spectroscopy requires funneling a thermal distribution across hundreds of quantum states into a single, well-defined state; this presents a computational challenge in expansive state spaces.
The team’s workflow uses operator learning to create a fast, differentiable surrogate model, the Fourier Neural Operator, capable of predicting population trajectories with a single forward pass given the molecule’s initial state and a Raman sideband pulse’s parameters. Compared to a reinforcement-learning baseline using the same discrete control space, the FNO-based approach approximately doubled the success rate of achieving the target state. This improvement extends beyond simply reaching the desired quantum state; the new method also reduced the number of pulses required and drastically shortened sequence-generation time.
The FNO cut sequence-generation time from around 10 hours to 10-20 minutes, NVIDIA says. These advancements, alongside partnerships with companies like Quantinuum, QuEra Computing, and IonQ, demonstrate NVIDIA’s commitment to accelerating quantum computing research and development.
FNO-SPMP Achieves 86.2% Success in Molecular State Preparation
Achieving an 86.2% success rate, the FNO-SPMP method represents a leap in molecular state preparation, exceeding previous benchmarks for polyatomic hydronium (H₃O⁺). This performance stems from a novel application of Fourier Neural Operators to design control sequences for molecules in exceptionally complex quantum states. The approach bypasses limitations of conventional optimization techniques, which struggle with the vastness of the hydronium molecule’s state space.
The significance of controlling hydronium extends beyond demonstrating a technical capability; its inversion transitions exhibit sensitivity to variations in fundamental physical constants, positioning it as a potential tool for probing physics beyond the Standard Model. Researchers targeted the molecule at 20 Kelvin, where it initially exists as a superposition across hundreds of rotational and hyperfine levels, requiring precise control to funnel it into a single, well-defined quantum state.
This is a considerable challenge in itself for polyatomic molecules. This combination allows for exploration of Hilbert spaces previously inaccessible to direct optimization, opening possibilities for other molecular ions and quantum platforms. “We can’t prompt our way to a control pulse sequence,” explains the UCLA NarangLab group, reinforcing the need for physics-informed models capable of generating effective control protocols. The resulting pulse sequences achieved a target-state population of 0.98, a result that demonstrates the potential of this approach to unlock new avenues in quantum control and fundamental physics research.
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
