PFDA-BS for Exascale-Quantum Hybrid Optimization

Predicting the ground-state energy of a molecule is a major challenge in computational chemistry, with direct applications in drug design, catalysis, and materials science.

The Variational Quantum Eigensolver combines quantum and classical computing, but it faces two practical limits: the combinatorial growth of the parameter space and the difficulty of using exascale resources efficiently. The objective is a generic optimizer that scales on hybrid exascale-quantum machines and does not require molecule-specific parameter tuning.

PFDA-BS was validated on the LUMI supercomputer with nine molecules covering a range from 3 to 560 parameters. Unlike existing approaches often limited to a few dozen accelerators, the method uses up to 8000 GPUs in parallel and reduces computation time from hours to minutes.

The optimizer reaches the chemical accuracy required in practice without gradient computations and without retuning parameters for each molecule.

For the first time, a generic black-box optimization algorithm can fully exploit exascale supercomputers for chemically relevant molecular simulations.

Because the method only uses energy values, it can be transferred without modification to hybrid workflows with quantum hardware once such machines become available. This makes it a practical bridge between current exascale systems and future exascale-quantum platforms.

Partners: Inria and Université de Lille.