Quantinuum, NVIDIA, and Pfizer trained a transformer to generate quantum circuits for imipramine calculations and ran two of them on Helios-1
Quantinuum, NVIDIA, and Pfizer trained a transformer to generate quantum circuits for imipramine calculations and ran two of them on Helios-1
On July 24, the team published a preprint on ADAPT-GQE. The model takes a description of the electronic structure problem for one imipramine configuration and generates a circuit for the calculation. The team then ran two 12-qubit circuits on the Helios-1 trapped-ion quantum computer.
Calculating a molecule's properties requires finding the lowest-energy state of its electrons. A quantum computer prepares this state by applying a sequence of operations to qubits. This sequence forms the circuit used for the chemistry calculation.
ADAPT-VQE builds the circuit one step at a time. It evaluates possible operations, selects the next operation according to how it changes the energy, and then optimizes the parameters of the entire circuit assembled so far. When the atoms change position, the search starts again. A flexible molecule has many such configurations.
ADAPT-GQE stores the results of this search in a training dataset. The transformer learned from circuits that ADAPT-VQE had built for different imipramine configurations. Given a description of a new configuration, it then generated 16 complete candidate circuits. The authors evaluated the energy of each candidate, and reinforcement learning steered the model toward candidates with lower energies.
For one imipramine configuration, the Gemma model generated 16 candidates in 1,6–3,7 seconds, whereas ADAPT-VQE required 2 333 to 25 341 seconds for a complete search. This comparison measures the time needed to prepare a circuit when the search is repeated for a new set of atomic positions.
The authors ran two 12-qubit circuits on the device. The compiler reduced the number of two-qubit operations from 1 440 to 244 in one circuit and from 1 248 to 283 in the other. Quantinuum then ran both circuits on the Helios-1 trapped-ion processor.