IBM, Mitsubishi Chemical and Keio train generative quantum circuits on 70 qubits

Started by BigDog_Fan, Today at 12:10 AM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Topic: IBM, Mitsubishi Chemical and Keio train generative quantum circuits on 70 qubits   Views(Read 13 times)
Active members in this topic:
BigDog_Fan(1)

BigDog_Fan

A transformer model has learned to design quantum circuits directly on real hardware. Quantum Computing Report says IBM Quantum, Mitsubishi Chemical, Keio University and NVIDIA trained a Generative Quantum Eigensolver, or GQE, on the 156 qubit IBM Heron processor ibm_kobe, using active spaces of 46 and 70 qubits

The target was organotin photoresist molecules used in extreme ultraviolet lithography for making chips. The main test case was methyltin trihydroxide and its hydrates, across 11 active space tiers

An autoregressive transformer generates the circuits, trained in two stages, first on a noiseless simulator and then directly on ibm_kobe. A penalty stopped it from collapsing onto the simple Hartree Fock state, and sample based quantum diagonalisation on GPUs finished the calculation

For 56 to 70 qubits, the circuits used 64 to 160 two qubit gates, against 4,248 to 6,952 for a standard LUCJ baseline. Across 46 to 70 qubits, 27 of 28 runs landed within 15 milliHartree of the DMRG reference, with a median gap of 4.5, and training took about 120 million shots across 600 jobs, with the work posted as an arXiv preprint

Cutting gate counts this dramatically matters because every gate on today's hardware adds noise. Using AI to design circuits rather than relying on fixed templates is a clever combination of the two fields. Chip making materials are also a very practical target, since better photoresists could help produce future processors

Can AI designed circuits make today's noisy quantum computers useful sooner? And is chip materials research a good early use for quantum chemistry?