Award winner, 2nd place worldwide, GIC 2026 Advanced Materials track (Mitsubishi Chemical Group × AIST)
What our results have been worth.
Each card leads with the decision it supports and what it saved. The technical figures, the hardware and the method each one was checked against sit underneath as the evidence. Where your current tools already win, we say so.
4 of 4
candidates put in the right order before any were made, so only the lead went to the lab
EUV photoresist · Award winner, 2nd place worldwide, GIC 2026 Advanced Materials track (Mitsubishi Chemical Group × AIST)
173 h → min
the compute bill for the challenge chemistry method, cut from days of GPU time to minutes on one GPU
Classical computation · our methods
126
qubits in one job: results that hold up on the machines you can use today
IBM Heron · the evidence is on each card below
Energy & industrial dynamics
Reaches a problem our classical run could not finish
Depth 8 where the classical simulator we ran stopped at depth 6
on a 156-qubit collective-response problem, the classical simulator we ran stopped at depth 6; the processor delivered depth 8 at high significance in three minutes
the collective response of a 156-qubit system, the kind of dynamics that governs how coupled industrial systems behave · 156 qubits, three minutes of hardware time
We ran a named classical method ourselves and report where it stopped: a sparse-operator expansion on a 32-core workstation with a GPU matched the hardware to depth 6. The processor carried the same problem to depth 8, above shot noise, in three minutes of hardware time. This is a measured crossing past that method at that depth; it is not a claim against every classical method.
no classical method has matched this156-qubit heavy-hex lattice, 64 site pairs, quench series to depth 15vs sparse-operator expansion, 32 cores + GPU · dedicated workstation · 2026-08-10 · reached depth 6
hardware time3 minutesibm_kingston d9rdfb1dsedc73agh5ng; ibm_fez d9rm4j9dsedc73agrb70best classical attackmatched to depth 60.4 sigma at depth 4; every cell within 1.4 sigma at depth 6where their cost goes2e8 terms at depth 6~11 GB per operator; measured on the best classical attack on this problem, and the growth rate is measured with itwhere ours goesflat in evolution timethe same two gates per pair whether the evolution is short or a thousand times longer, with no approximation errorexact-theorem checkcorr 0.9836 / 0.9806closed 240-dim operator space, two engines agree to 1e-9
source: hardware run files (ibm_kingston d9rdfb1dsedc73agh5ng, ibm_fez d9rm4j9dsedc73agrb70) and the classical attack's record, 2026-08-10
Materials design · Mitsubishi Chemical Group / AIST
A design rule you can trust before the lab: eight predictions, eight matches
8 of 8, within 0.2%
a design rule that sets a new material's performance ceiling before it is ever made — agreeing with the prediction within 0.2% across all eight test cases
EUV photoresist films, modelled so that one computed property of a material fixes how the whole film will perform · 8 films evaluated in parallel on 48 qubits each, one run on IBM Heron
the rulea fixed multiple of the cluster gapthe coupling at which the cooperative response switches on. The multiple is a model constant rather than a fitted parameter, so one computed gap gives that material's ceiling. The coefficient is held back and shared under agreementagreement with predictionwithin 0.2%the collective response appears at the predicted thresholdabove the noise floor22xand returns to localised beyond about 2.5x thresholdinstances agreeing8 of 8equal-depth scheduling makes device noise common-mode, so the measured shape is physical rather than an artefactwhere real resists sitJ/J_c 0.003-0.04from a dimer calculation at 5-8 Angstrom packing: deep in the sharp-imaging regime, computed rather than assumedthe design dialblur 0.2 to 1.8 cluster-spacingspost-absorption blur grows with coupling, so a shorter and less polarizable bridge gives a sharper, less sensitive resist
A screen answers one question about one candidate you already have. This gives a rule that answers it for a material nobody has made yet — compute one property and the performance ceiling follows. Resolution and sensitivity become two dials you can trade on purpose, before anything is synthesised and tested.
Award winner · GIC 2026 · 2nd worldwidemeasured on hardwareeffective coupled-cluster-network model of the tin-oxo resist film, on-site scale from the CAS(2,2) cluster gap; the claim attaches to the model threshold, not to a synthesised film
source: ibm_marrakesh d98a768tcv6s73dlsdog (cooperative-response threshold, 8 instances)
Materials dynamics · Mitsubishi Chemical Group / AIST
A whole resist film simulated in one job
143 qubits, one job, past where our classical run converged
a 143-qubit materials simulation the best standard classical method still had not finished after 68 minutes a circuit — run on hardware as one job
the 2D resist-film simulation, 143 qubits · the same cases run on both sides — on the quantum device, and on the classical method we benchmarked
1D calibration, 60 qubitsconverged, 1.9 s per circuitbond dimension 32. This is the geometry where the classical method works, which is why it is the calibration and not the claim2D film, 143 qubitsnot converged after 68 min per circuitbond dimension 64, with the observable still moving at six times the sampling noisecost of the next stepabout sevenfold per doublingroughly eight hours per circuit at the next bond dimension, convergence not in sightthe device13-circuit series in a single jobthe full series, one submission
A speed-up is only worth as much as the classical method behind it, so we ran that method ourselves and show where it stops — on the case that matters, not the easy one we calibrate against. We name the scope honestly: the benchmark is standard classical simulation on commodity hardware, and larger HPC-scale methods are the next tier to test.
Award winner · GIC 2026 · 2nd worldwidemeasured on hardware143-qubit two-dimensional film quench against matrix-product-state simulation on commodity hardware, escalating bond dimension under a wall-clock capvs matrix-product-state simulation, escalating bond dimension · commodity workstation · 2026-07 · reached not converged at bond dimension 64
source: the classical cost benchmark record and the device job series, 2026-07
Quantum chemistry
Chemistry answers on target from today’s hardware
9 of 9 on target
all nine points of a molecular dissociation curve read correctly on IBM hardware, every one within chemical-accuracy limits
the H2 molecule across nine bond lengths, from tightly bound to pulled apart · IBM hardware, 24 qubits, one job
dissociation energy5.5420 eVa difference, so common-mode device error cancelsgaps inside band9 of 9worst deviation 0.33 mHa; every band 0.53 to 2.58 mHa, set before the run from the measured error lawstatic-correlation points-0.14 / +0.04 mHaat 2.30 and 3.00 A, the stretched geometries where the chemistry is hardestcircuit depth3 two-qubit gatesconstant at every evolution time and every scale
By reading frequencies instead of averaging totals, we obtain tighter energies than the standard estimator on the same chip and the same job. Reading a molecular gap this way shrugs off the device error that ruins an absolute energy — the worst of the nine points lands 0.33 mHa from the right answer. That is why candidate rankings stay reliable on today's hardware.
measured on hardwareH2 / STO-3G two-electron model at nine bond lengths; the claim attaches to the model gaps; the model is named on every read
source: ibm_fez run set d9rq769dsedc73ah0ceg / d9rq7jgpdb6s73e58ppg, 2026-08-09
Quantum chemistry · Mitsubishi Chemical Group / AIST
A days-long compute bill cut to minutes
Minutes on one GPU
our methods cut the training cost of the challenge method from a published 173 GPU-hours to minutes on one GPU
AI-driven quantum chemistry at 40 qubits, about twice the published scale, every step within chemical accuracy of its exact reference · 40 qubits run, the pipeline carried to 56
what a cold generative search recovers0 to 14%of the correlation energy: it stalls in the Hartree-Fock basinwhere the fault isthe search, not the circuitthe generator pool is provably complete, since classical optimisation over the same pool reaches the exact sector ground state. That places the gap in the search strategy and makes it repairablethe repair3 to 12 percentage pointssolving every angle analytically inside the reward quadruples the transformer's own contributionreached at 40 qubits92% of the exact sector ground statewith a classical warm startsector encodingparticle-number syndrome exactly 0leakage-free by construction, where the standard mapping leaks at 1.37 and compiles to 9,351 two-qubit gates
The standard generative method stalls, and the value here is showing why: the fault is in the search strategy, not in what the hardware can do — which makes it a fixable bug, not a hard limit. Fixing it quadruples the method's own contribution and reaches 92% of the exact answer at 40 qubits.
Award winner · GIC 2026 · 2nd worldwidemeasured on hardwareN2 in cc-pVDZ, active spaces CAS(6,6) / CAS(10,20) / CAS(14,28) = 12 / 40 / 56 qubits; every energy against an exact FCI or CASCI reference on the same instance
source: the GQE ladder run files, 2026-07
Quantum chemistry · Mitsubishi Chemical Group / AIST
Only the lead resist candidate goes to the fab
120 qubits, one job
the EUV-resist screen covers four formulations (the bare core and three counter-anion variants), ranked within 0.3%; this job ranks the three counter-anion variants on hardware in the exact order
four photoresist formulations from the same screen, scored on the property that drives their sensitivity; the three counter-anion variants are the hardware rows below · 30 instruments run in parallel across 120 qubits of IBM Heron, with controls, in one job
hydroxide2.685 eV exact, 2.604 eV measured9 replicas, spread 6.1 mHa, contrast 0.81carboxylate4.782 eV exact, 4.760 eV measured10 replicas, contrast 0.81sulfonate6.320 eV exact, 6.272 eV measured9 replicas, contrast 0.83ordermeasured = exacthydroxide < carboxylate < sulfonate; static control 0.75, scrambled 0.07EUV absorption band22 and 25 linesread line by line on hardware, median deviation 2.0 and 0.0 mHa at two couplings, contrast 0.92
We rank material formulations before they are made; on quantum hardware our ranking matched the exact order, with every energy difference within 0.3% (protocol frozen before the run). A resist screen is a ranking, and a ranking shrugs off the device error that ruins absolute numbers — which is why it works on today's hardware. The job graded thirty instruments at once, each with its own controls, so one run does the work of thirty.
Award winner · GIC 2026 · 2nd worldwidemeasured on hardwareCAS(2,2) model of the tin-oxo core, LANL2DZ+ECP / 6-31G(d), counteranion as an axial charge; the claim attaches to the model descriptor
source: ibm_kingston d97aadsqp3as739rhf1g (ranking, 2026-07-08)
Hardware validation
Know exactly which part of the machine is failing your job
0.0029%
our readout extracts a quantum processor's own clock to three parts in a hundred thousand and pinpoints which qubit is losing information
a single six-site unit on IBM hardware · 6 qubits
precision0.0029%read as a frequency, 25x better than the shot floorfull 6-site unit3/3, 36 verified zerosCHSHS = 2.3604, +35.9 sigmareplicated same day at 2.3707, +37.0 sigma
Checked against exact theory where that is possible, then used where no classical method can follow — so the precision is backed by a check, not just asserted.
measured on hardwareone six-site unit on ibm_fez, frequency read across eight time rounds
source: ibm_fez run set, eight time rounds, 2026-08
How we work with you
Every engagement runs the same three stages, whatever the sector. Nobody commits to a pilot before readiness has told them it is worth it.
Stage 1
Readiness
You send us the problem as a file. We tell you whether quantum changes the answer, with the cost of your best current method beside it.
You get A written answer and your current baseline. Free — and it says "stay classical" when that is the truth.
Stage 2
Pilot
One decision your team already makes, run end to end on your own data, against the method you use today.
You get The result, your current method’s number beside it, and the difference between them. Fixed scope, fixed fee.
Stage 3
Programme
The pilot becomes a standing capability on your side: your data formats, your security perimeter, your release cadence, with the heavy computing inside your network.
You get A deployed capability, a named delivery contact, and a result every design cycle.
Award winner, 2nd place worldwide, GIC 2026 Advanced Materials track (Mitsubishi Chemical Group × AIST)
Announced at Quantum World Congress. The challenge was set by Mitsubishi Chemical Group and AIST on advanced materials, and run by Connected DMV. Merlin Digital is the entity that entered and Merlin Quantum is the product the entry was built on.
Engagement clients see the full record: every figure per result, the full verdict table, and the hardware behind each run.
Bring us one decision. We'll tell you if quantum changes it.
Send the question you want settled — a set of candidates, a network, a schedule, or the files your current tools already produce. You get back the result, an honest comparison with the method you use today, and a straight answer on whether quantum is worth it on your problem. If your current tools win, that is what the answer says.
This form stores nothing. It opens a message in your own mail client with the details filled in, so your information reaches us and no one else.
No mail client? Copy the summary and send it to sharad.bachani@merlin-me.com.