AI Learns to Build Quantum Circuits 1,000× Faster for Drug Design

Researchers from Quantinuum, NVIDIA, and Pfizer have combined transformer-based language models with reinforcement learning to automate the generation of quantum circuits for molecular simulations. The work addresses a longstanding bottleneck in quantum chemistry: before a quantum computer can estimate the energy of a molecule, it must first prepare the quantum system in a state that closely approximates the molecule's ground state, or lowest-energy configuration. Producing those circuits typically requires repeated optimization that becomes increasingly expensive as molecules grow larger.

The team's AI system reduced circuit-generation time by three to four orders of magnitude compared with ADAPT-VQE (Adaptive Derivative-Assembled Problem-Tailored Variational Quantum Eigensolver), a leading conventional quantum chemistry algorithm, while matching or exceeding its accuracy. Rather than constructing circuits through repeated optimization, the trained models generate complete circuits in a single inference step. The innovation frames quantum circuit generation as a language problem: instead of generating sentences, transformer models generate sequences of quantum operations needed to prepare molecular ground states.

The researchers trained two different transformer architectures. One was a relatively compact 325 million-parameter model trained from scratch; the other was a much larger 12 billion-parameter pretrained language model that was adapted for quantum circuit generation. Both models received information describing the molecular Hamiltonian—the mathematical representation of a molecule's electronic interactions—and generated corresponding quantum circuits. For the smaller model, the team added a reinforcement learning stage after supervised training. Rather than simply reproducing circuits from the training data, the model received rewards for generating circuits with lower calculated energies, enabling the AI to surpass the quality of the original training examples in many cases.

The team selected imipramine, a tricyclic antidepressant, as its primary test system because of its structural complexity and pharmaceutical relevance. Imipramine has multiple reactive sites and can adopt many different three-dimensional shapes, known as conformers. Predicting the electronic properties of these different configurations is important in pharmaceutical development, particularly when studying drug stability and degradation pathways. The researchers generated roughly 13,000 to 15,700 circuits representing a broad range of molecular geometries and circuit complexities, using ADAPT-VQE to produce optimized quantum circuits across 12-, 14-, and 16-qubit active-space representations.

Both transformer models reproduced the performance of ADAPT-VQE for easier molecular problems, but performance declined after supervised training alone on more complex quantum circuits. Applying reinforcement learning substantially improved the smaller model's performance across every dataset. The optimization process frequently produced circuits with lower energies than those found in the original training data, suggesting the AI discovered improved solutions during training. The researchers also observed different behavior between the two model types: the smaller transformer generated more consistent, narrowly distributed solutions, while the larger pretrained language model produced a broader range of candidate circuits.

Beyond simulation, the team executed representative AI-generated circuits on Quantinuum's Helios trapped-ion quantum processor, demonstrating that circuits produced entirely by the AI framework can operate on existing quantum hardware. The researchers described this as a milestone because it combines AI-generated circuit synthesis with execution on commercial quantum hardware for a chemically relevant molecular system. Previous work in this area has generally focused either on simulations or on relatively simple benchmark problems rather than realistic pharmaceutical molecules. The researchers note several limitations, including that the models were trained separately for each qubit count, meaning they cannot yet generalize across different active-space sizes, and validation has focused on a single pharmaceutical molecule. Broader testing across larger and more chemically diverse molecules will be needed.