An AI learns from a quantum chip to propose shorter chemistry circuits

Learning which experiments to run

The study, available as an arXiv preprint and not established here as peer-reviewed, tackles a practical problem in quantum chemistry: preparing a useful electronic state before hardware noise overwhelms it.

Its circuit designer is a classical transformer, not a quantum neural network. Given descriptors of a molecule, it generates a sequence of operations for a quantum processor. Measurements from those circuits feed a hybrid calculation called sample-based quantum diagonalization, or SQD.

In SQD, the quantum computer does not produce the final molecular energy directly. It samples electronic configurations—possible arrangements of electrons among the selected orbitals. A classical computer uses those samples to construct a restricted space of configurations and solves the molecular energy problem within it.

That distinction matters. The final answer belongs to the combined quantum-and-classical workflow, not simply to the state prepared on the chip.

From simulation to hardware feedback

The researchers first trained their generators on a noiseless simulator using molecules represented by at most 32 qubits. They then continued training with IBM’s Heron-generation processor, ibm_kobe, providing measurement feedback for transition-metal complexes at 46 and 70 qubits.

Four main training runs combined two molecular-conditioning approaches with two circuit lengths. One conditioning approach used numerical molecular descriptors; the other added structural features from a frozen, pretrained model.

Together, these hardware-training runs used 120 million measurement shots and 648 minutes of billed quantum-processing time. Classical computation and device queues added substantially more elapsed time.

Training also required approximate classical reference energies. Those references controlled whether a batch could update the model and whether two auxiliary penalties applied. The authors did not test a reference-free training protocol, although evaluation on the target molecules required no reference energy to select circuits.

One essential penalty discouraged a tempting shortcut: generating circuits whose valid measurements merely reproduced the starting Hartree–Fock configuration, a simplified electronic reference.

Trying different chemistry

The trained policies were then applied without further training to methyltin trihydroxide and its one- and two-water adducts. These are minimal models relevant to organotin photoresists for extreme-ultraviolet lithography, not simulations of a complete photoresist or its response to light.

The targets covered eleven selected electronic spaces, spanning 32–70 qubits. For each policy–target combination, the generator proposed ten circuits, each measured 100,000 times. A relatively inexpensive classical calculation ranked the candidates; more extensive configuration recovery then reused the stored measurements without further quantum-processor calls.

For the main training recipe, 43 of 44 combinations produced a reported recovered best-circuit result. Their median deviation from density-matrix renormalization group, or DMRG, reference energies was 2.8 millihartree. A millihartree is one-thousandth of the hartree energy unit. Forty-two results were within 15 millihartree—useful agreement, but not equivalent to uniformly reaching the approximately 1.6-millihartree threshold commonly called chemical accuracy.

One combination was excluded after producing only one circuit with valid measurements; a second sampling attempt still gave a large recovered error. In all eight comparisons with an untrained generator, the trained policy was more accurate. Those controls covered only three target systems.

Short circuits, substantial supporting work

At 56–70 qubits, selected circuits used 64–160 two-qubit gates, between 34 and 72 times fewer than the study’s transpiled local unitary cluster Jastrow, or LUCJ, comparator.

But selection is doing important work. Many other generated circuits were deep, and the typical candidate was not necessarily cheaper than LUCJ. Sampling ten candidates consumed about eleven times the quantum-processing time per job of sampling the single LUCJ circuit. Final classical recovery dominated evaluation wall time.

The physical interpretation is also unusual. On three smaller targets examined exactly, selected ideal circuits remained close to the Hartree–Fock starting state. Hardware measurements contained substantially more excited configurations; the authors conclude that device noise supplies excitation content exploited by recovery. That observation does not establish noise as generally beneficial.

What the experiment establishes

The evidence supports hardware-in-the-loop circuit-generator training at 70 qubits, not a demonstrated computational advantage over classical chemistry.

Each training lineage used one seed, most circuits were sampled only once, and the main policy narrowed to a very limited repertoire during training. Small-target accuracy was helped by recovery spaces large enough to accommodate every possible spin-sector string. Moreover, although the tin targets were absent from policy training, an earlier tin-target result informed one reward weight, limiting claims of untouched generalization.

The achievement is therefore specific: a learned proposal-and-selection workflow can produce short, useful circuits on noisy hardware. Whether its training, sampling and classical-recovery costs pay off on harder chemistry remains unresolved.