204
Part Five · Convergence

Hybrid pipelines: classical ML, quantum subroutines, and HPC working together

No quantum computer will replace a drug-discovery or materials pipeline. The realistic picture, already being built, is a quantum processor inside a supercomputer, called for one narrow task while classical machine learning, density functional theory and molecular dynamics do everything else. In 2025 a team from IBM and RIKEN ran such a workflow on iron–sulfur clusters, the same kind of active sites as in FeMoco, with up to 77 qubits on IBM’s Heron processor and Japan’s Fugaku supercomputer doing the heavy classical lifting. This chapter shows how the pieces fit, and what the 2025 result did and did not show.

On the plate: a pipeline you assemble, with its costs.

Funnels

Discovery pipelines are funnels: cheap methods screen millions of candidates; progressively more expensive and accurate methods examine fewer. Machine learning trained on past data scores a million molecules in hours; docking or semi-empirical methods trim to thousands; DFT examines hundreds; correlated methods, or one day a quantum subroutine, examine a handful of hard cases; free-energy simulations rank the finalists; and the lab tests a few. A quantum step makes sense only where it is the accuracy bottleneck for the few cases that reach it. Established as practice.

Build a pipeline

Include stages: ML pre-screen ; DFT refinement ; quantum active-space step ; free-energy ranking . The plate tracks candidates and compute cost through each stage from a library of a million molecules, with illustrative costs per molecule. Notice that without the ML screen, every later stage is swamped; that the quantum step’s cost is small in total because few candidates reach it; and that its value depends on whether the candidates at that stage have hard electronic structure. Established as a design pattern; numbers are illustrative.

Sample-based quantum diagonalization

In 2025 Javier Robledo-Moreno, Antonio Mezzacapo and colleagues at IBM, with RIKEN collaborators, published in Science Advances a method called sample-based quantum diagonalization (SQD): the quantum processor samples electronic configurations from a circuit approximating the ground state, and a supercomputer diagonalizes the Hamiltonian in the subspace spanned by the sampled configurations, correcting noise along the way. They treated [2Fe-2S] and [4Fe-4S] cluster models with up to 77 qubits. It is the largest quantum-centric chemistry workflow to date. A co-author was quoted as saying the result does not yet show that it beats any classical algorithm, and comparisons against selected CI and DMRG on the same problems remain under discussion. Established as a workflow; Debated for advantage.

What you now have

Discovery funnels, where a quantum subroutine would sit, how stage ordering controls cost, the 2025 sample-based quantum diagonalization of iron–sulfur clusters, and why benchmarks against the best classical methods are the open question.

· · ·

Three questions to carry with you.

Why put ML first in a funnel?tap
It is cheapest per molecule and removes most candidates before expensive stages.
What does SQD use the quantum processor for?tap
Sampling important electronic configurations; classical diagonalization does the rest.
What remains contested about the 2025 result?tap
Whether it outperforms the best classical methods, such as selected CI and DMRG.

And one to work out. An ML screen keeps the top 1% of a million molecules. How many go on?

≈

Sources: J. Robledo-Moreno et al., Sci. Adv. 11, eadu9991 (2025). R. Santagati et al., Nat. Phys. 20, 549 (2024). Stage costs in the plate are illustrative orders of magnitude.

Next · Chapter 205
Reality check: what has been demonstrated on hardware versus proposed

A table with three columns: demonstrated, estimated, hoped.

Turn the page