FLiBe Fusion Blanket Quantum Chemistry

Om Labs Quantum Research OS — Research Recreation

FLiBe Fusion Blanket Chemistry

ORNL × Cleveland Clinic × IBM Quantum

A full recreation of the first known quantum computation on fusion reactor materials — from FLiBe cluster chemistry and embedded wave function fragments to LUCJ circuits, ext-SQD simulation, and a real IBM Fez hardware stamp.

Status: Phases 0–9 Complete

Method: EWF + ext-SQD

Hardware: ibm_fez ● 38.2s

Paper: arXiv:2606.30402

Context

Why Tritium Chemistry on a Quantum Computer?

Fusion energy needs tritium. Tritium is scarce. FLiBe molten salt must breed it and give it back. Predicting that chemistry accurately requires correlated quantum calculations.

21–23

atoms per cluster snapshot

FLiBe conformations studied

qubits, largest LUCJ circuit

0.3

kcal/mol MAD, solver vs FCI

110

kcal/mol embedding bias

⚛

The Physics Goal

Fusion reactors burn deuterium + tritium. Tritium is rare — the whole planet's supply fits in a small container. Blankets of FLiBe molten salt must breed new tritium from lithium-neutron reactions and then recover it controllably. This requires understanding how tightly tritium binds in the salt at the electronic level.

🔬

What ORNL/IBM Actually Computed

Not a full reactor simulation. They took 9 frozen cluster snapshots (21–23 atoms each) from AIMD molecular dynamics. They split each into atom-centered fragments via EWF embedding, solved small ones with FCI, large ones with quantum-powered ext-SQD on IBM Heron r3, then reassembled binding energies. This is fragment quantum chemistry, not reactor engineering.

👉

The Key Finding

Quantum fragment solves (ext-SQD) agreed with classical FCI within ~0.7 kcal/mol MAD ~0.3 kcal/mol on conformational energy differences. But the embedding / how you cut the cluster dominated binding energy error — ~110 kcal/mol offset vs full-molecule methods. The quantum solver is not the bottleneck. Smarter embedding is.

📈

Our Binding Observable

Every number in this project traces back to one equation:

Eₛ𝕫ₙ𝔡 = E(FLiBeTF) − E(FLiBeF⁻)

Tritium leaves as bare T⁺ with zero electronic energy (paper definition). Strongly negative = T energetically prefers staying in the salt.

Learning Story

The Salt Keepers — 7 Episodes

Click any episode to expand the intuition. Each episode pairs with a technical lesson.

Scientific Method

The EWF + ext-SQD Pipeline

From frozen cluster snapshot to fragment energy and back to binding energetics.

AIMD/MLFF Cluster Snapshots

Nine FLiBe conformations (21–23 atoms: Li₆Be₃F₁₂, anion, +T). Born-Oppenheimer: nuclei frozen for each electronic solve. Our recreation uses chemically faithful proxy clusters.

Phase 2 done

RHF Mean-Field + Integrals

PySCF density-fitted RHF in 6-31+G(d). Builds orbitals and one/two-electron integrals hₖᵒ and gₖᵒᵣₘ. 27/27 clusters converge. This is the foundation everything builds on.

Phase 3 done

Vayesta EWF Fragmentation

IAO localization → DMET Schmidt bath → MP2 bath expansion (η=10⁻⁵). Each atom gets a fragment Hamiltonian H𝔡 with M orbitals. F-centered fragments balloon to M~27–33 due to polarizable lone pairs.

Phase 4 done

Fragment Split Rule

M < 13: exact FCI (classical, gold standard). M ≥ 13: ext-SQD path (LUCJ sampling + classical diagonalization). This hybrid is the paper's core innovation.

Phase 4-5

Classical FCI / CCSD Ladder

Classical teacher: HF → MP2 → CCSD → FCI/TCI. CCSD T₁,T₂ amplitudes seed LUCJ. FCI is the gold reference for small fragments. T₁ diagnostic ~0.014–0.016 confirms single-reference character.

Phase 5 done

LUCJ + ext-SQD Simulation

ffsim LUCJ (nᵢᵒᵖₘ=1) seeded from CCSD → Aer Sampler bitstrings → qiskit-addon-sqd configuration recovery → compact CI subspace diagonalization → subspace extension (singles) → fragment energy + 2-RDM.

108 fragments solved

Energy Reassembly

Partitioned cumulant: fragment 1- and 2-RDMs → correlation energy on top of global HF → EWF-FCI+ext-SQD total → ΔE conformational + Eₛ𝕫ₙ𝔡 binding.

Phase 7 done

IBM Quantum Hardware Stamp

One gated Sampler job on ibm_fez. 26-qubit LUCJ. 4096 shots. 38.2s RUNNING. Job d9l0333jf64c739ir200. Fragment energy within 5.6 kcal/mol of noiseless sim.

⚡ Phase 8 complete

Project Execution

10 Phases — All Complete

Every phase writes artifacts into the Research OS knowledge trail. Nothing is skipped.

Quantum Hardware

IBM Fez — The Real Stamp

One gated Sampler job on a production IBM Quantum system. Budgeted ≤120s RUNNING. Executed clean.

⚡ ibm_fez SamplerV2 LUCJ n_reps=1

Job ID

d9l0333jf64c739ir200

ibm_fez backend

Qubits

LUCJ circuit (M=13 fragment)

Shots

4,096

vs 10⁵–10⁶ paper campaign

Transpile Depth

752

1,170 two-qubit gates

Wall Time

64.3s

RUNNING: 38.2s ✔

Usage

billed QPU seconds

−52.117

Eₘⴀ𝔔 (sim) Ha

→

−52.126

Eₘⴀ𝔔 (hw) Ha

−5.6

Δ kcal/mol (hw − sim)

✅

Kill switch: not fired

Budget policy respected

⚠ Attempt 1 — Queue Timeout (Documented)

Backend: ibm_marrakesh | Job: d9kuplqbr2fc73e7u0g0 | Queued ~30 min → cancelled by queue ceiling (1800s). QPU burn: 0. Lesson: least-pending backend beats marketing qubit count. Switched to ibm_fez → success. Both attempts archived in data/.

Analysis

Results & Scoreboard

What we matched, what we honestly could not, and why the gap is instructive.

RHF Binding Energetics (9 pairs)

Metric	Our Value	Paper Band
Mean Eₛ𝕫ₙ𝔡	−286 kcal/mol	~−134 to −380
Range	−407 … −122	Strongly negative
Direction	✓ Matches	Strongly bound T
SQD fragments (sim)	108	All M≥13
Solver Δ (hw vs sim)	5.6 kcal/mol	Absolute; cancels in ΔE

The Key Science Lesson

⚠

Embedding Error >> Solver Error

Quantum fragment solve (ext-SQD) vs FCI: ≤0.7 kcal/mol MAD. Embedded family vs uncut full-molecule methods: ~110 kcal/mol offset on binding. The QPU is not the bottleneck. How you cut the cluster is. This is the paper's most important message — and we reproduced the qualitative lesson.

Claim Scoreboard

⚛

Same chemistry (FLiBe / TF / anion stoichiometry)

✓ Yes

🔬

Same method family: EWF + LUCJ + SQD

✓ Yes

📚

Same basis set: 6-31+G(d)

✓ Yes

〈

Same M-split rule (FCI <13; SQD ≥13)

✓ Yes

📊

Qualitative T strongly bound direction

✓ Yes

⚡

Real IBM Sampler path (ibm_fez)

✓ Yes

📄

Exact paper AIMD cluster geometries

✗ No (proxies)

🛠

Full M~33 / 66-qubit coverage

✗ No (413 large skipped)

💻

Paper shot campaign (10⁵–10⁶)

✗ No (4096 shots, 1 job)

⚖

Full EWF Eₛ𝕫ₙ𝔡 reconstruction tables

~ Partial (RHF binding)

Learning Roadmap

7 Technical Lessons

Click any lesson to expand. These unlock the full pipeline step by step.

Technology

Stack & Tools

PySCF + Vayesta + ffsim + qiskit-addon-sqd. No silent swaps for paper methods.

📥

What this is NOT

Not full reactor MHD. Not DFT-only. Not blind VQE on the whole 23-atom cluster. Not claiming paper's 10⁶-shot campaign. Not skipping embedding to run a toy H₂.

📝

Research OS Artifacts

Every phase writes JSON/Parquet/MD into the knowledge trail. Manifest, Hamiltonians, SQD summaries, analysis reports, hardware results, KG nodes — all logged.

🛠

Hardware Discipline

Classical + Aer for all development. IBM Quantum Runtime only as final validation. ≤120s QPU budget. Kill switch armed. Dry-run transpile before any real job. Auth before burn.

Lab Journal

10 Lessons Tattooed on My Brain

First-person learner voice. After Phases 0–8. Still slightly high on ibm_fez.