Metal-organic frameworks and the quantum screening pipeline

Metal-organic frameworks are the most promising scaffold we have for direct-air carbon capture, and they are also the hardest class of material to screen properly on classical hardware. A single MOF unit cell can carry hundreds of atoms, transition-metal nodes with strongly correlated d-electrons, and a binding-site chemistry that depends on conformational states DFT routinely gets wrong. The screening problem is real, the chemistry is unforgiving, and brute-force enumeration of the hypothetical MOF space — well into the hundreds of thousands of candidates and counting — is why this is the workload I keep coming back to when I think about what a 100-qubit transmon machine in Tipperary should actually do in its first year of operation.

Why MOFs are a quantum-shaped problem

A MOF is a coordination polymer: metal nodes (often Cu, Zn, Mg, Fe, or rare earths) connected by organic linkers into porous crystalline lattices. The pores are where CO₂ binds. The binding energetics depend on the electronic structure at the metal node, the polarisability of the linker, and the cooperative effects of neighbouring sites. That is three layers of correlated-electron behaviour stacked on top of each other.

Classical computational chemistry handles this badly. Density Functional Theory is the workhorse, but DFT functionals are calibrated against datasets that under-represent open-shell transition metals and dispersion-dominated guest binding. Coupled-cluster methods (CCSD(T), the so-called "gold standard") scale as O(N⁷) with system size and become unworkable past a few dozen heavy atoms. So the field has been forced into a compromise: use cheap DFT to screen millions of structures, accept that the ranking is noisy, and synthesise the top candidates anyway. Plenty of promising MOFs never get found because they sit below the noise floor.

This is exactly the regime where quantum chemistry on quantum hardware has a genuine, defensible advantage. Variational Quantum Eigensolver (VQE) and its successors don't replace DFT — they replace the expensive correlated-electron calculation at the active site, after the classical pipeline has narrowed the field.

The screening pipeline, end to end

The pipeline I am building toward looks like this, and none of it is exotic — most of it is well-established classical workflow with a quantum kernel inserted at the bottleneck:

• Stage 1 — Structural enumeration. Generate or import candidate MOFs from existing databases (CoRE MOF, hMOF, QMOF) and apply geometric filters: pore size, surface area, accessible volume.
• Stage 2 — Classical DFT pre-screen. Plane-wave DFT (VASP, Quantum ESPRESSO) to compute approximate binding energies. Rank, keep the top few percent.
• Stage 3 — Active-site extraction. Cluster models or embedded-fragment methods to isolate the metal node plus first coordination sphere. This is the chemically meaningful sub-system, typically 20–60 atoms.
• Stage 4 — Quantum kernel. Map the active-site Hamiltonian onto qubits via Jordan-Wigner or Bravyi-Kitaev, run VQE or quantum-subspace expansion, recover the correlation energy DFT missed.
• Stage 5 — Re-rank and synthesise. The quantum-corrected ranking points the lab at MOFs the classical pipeline would have buried.

Stage 4 is where Ireland Quantum 100 earns its keep. Everything else runs on existing CPU/GPU infrastructure and is well-understood.

What 100 physical qubits actually buys you

Let me be careful here. One hundred superconducting transmon qubits arranged on a heavy-hex topology, operating at sub-15 mK in a dilution refrigerator, without surface-code error correction, is a noisy intermediate-scale quantum (NISQ) device. It is not running Shor's algorithm and it is not factoring anything interesting. What it can do, with good calibration and short-depth circuits, is small active-space chemistry.

An active space of, say, 20 spin-orbitals maps to 20 qubits under Jordan-Wigner. That leaves headroom for ancillas, error-mitigation overhead (zero-noise extrapolation, probabilistic error cancellation), and parallel circuit execution across the device. A 20-orbital active space is genuinely useful for a metal-node fragment with one or two transition-metal centres and their immediate ligand environment. It is not the whole MOF — it is the chemically active piece, embedded in a classical mean-field bath.

The honest framing: in the NISQ era, what we are doing is producing reference data for a fragment-embedding workflow, not solving the full crystal. That reference data is still valuable, because the alternative — CCSD(T) on the same fragment — is either intractable or under-converged for open-shell transition-metal systems.

The error-mitigation reality

VQE on hardware lives or dies on three things: circuit depth, measurement count, and the quality of the ansatz. On heavy-hex transmon topology, two-qubit gates are the dominant error source. A hardware-efficient ansatz with O(10) layers on 20 qubits is feasible; a chemistry-inspired UCCSD ansatz at the same scale is not, because the gate count blows past the coherence budget.

So the practical path is some combination of:

• ADAPT-VQE — grow the ansatz iteratively, only adding operators that materially reduce energy, keeping circuit depth disciplined.
• Qubit tapering via Z₂ symmetries — remove qubits that are constrained by particle-number or spin conservation.
• Zero-noise extrapolation — run the same circuit at amplified noise levels, extrapolate back to the noiseless limit.
• Classical shadows for efficient measurement of multiple observables from the same state preparation.

None of this is invented for us. It is the working toolkit that Qiskit, PennyLane, and Cirq all support, and it is what any honest superconducting platform offers. The job of a sovereign 100-qubit machine is not to invent new algorithms — it is to give Irish and European chemistry teams a reliable, low-latency, queue-free environment to run them on real workloads. Read more about the national quantum platform we are standing up.

Why carbon capture and not something else

I get asked this often. The answer is workload fit, not fashion. Carbon-capture MOF screening hits four criteria simultaneously, and I have not found another workload that hits all four:

• The classical pipeline already exists. We are not asking chemists to invent new methods — we are slotting a quantum kernel into a workflow they already run.
• The bottleneck is genuinely correlation energy at transition-metal sites. This is the kind of problem quantum hardware is theoretically suited to.
• The active-space size matches NISQ-era device capability. Twenty to forty qubits of usable active space, which is what we will have, is exactly the right scale.
• The downstream impact is measurable. A better MOF means a cheaper tonne of CO₂ removed, which feeds directly into the offset stack we run at IMPT. If a screening run surfaces a candidate that a synthesis lab can validate, that closes the loop in months, not decades.

Compare this to, say, drug discovery, where the protein active sites are larger, the classical pipeline is more entrenched, and the regulatory tail is a decade long. MOFs are short-loop. That matters for a first-year machine that needs to demonstrate value to the people funding it.

The integration with chemistry SDKs

Operationally, a chemistry team should be able to take a fragment Hamiltonian generated by PySCF or Psi4, hand it to OpenFermion or Qiskit Nature for the qubit mapping, and submit the resulting circuit through OpenQASM 3 to the device. We are not asking anyone to write quantum assembly. The submission layer looks like a job queue, the calibration data is published per-shift, and the noise characterisation is exposed so error-mitigation passes can be tuned per run.

If you want to see how the workload maps to the broader candidate pipeline, I have written separately about the climate workloads cohort and how MOFs sit alongside battery materials and photovoltaic discovery. They share infrastructure but have very different active-space profiles.

Where to start this week

If you are a research group, a chemistry department, or an industrial team thinking about MOF screening with a quantum kernel: the practical move this week is to take one of your existing top candidates from a DFT pre-screen and write it up as a fragment-embedding problem. Identify the active site, count the spin-orbitals, work out what your active space looks like at 20, 30, 40 qubits. That exercise alone tells you whether your problem fits the hardware we are bringing online — and it costs nothing but an afternoon. When the cryostat is cold and first-light is done, the teams that have already done that homework are the ones that will be running real chemistry first. The rest will be queueing.