Internal Record

CGA GPU Annealer on a Hard CO2RR Catalyst-Portfolio QUBO

Quantum-Inspired · CGA GPU annealer · 100 variables · 2026-08-17

GPU-accelerated simulated annealing benchmark on a hard CO2RR catalyst portfolio QUBO instance.

QUBO objective value

5.3 dimensionless energy→4.8 dimensionless energy-0.6 dimensionless energy

Gap to certified optimum

953.9 percent→840.2 percent-113.7 percent

Problem

Select a diverse, high-value batch of CO2 electroreduction (CO2RR -> CO) catalyst candidates from a predicted pool under a fixed validation budget. The selection trades predicted catalytic utility against pairwise chemical redundancy and enforces an exactly-K cardinality constraint.

Objective: Minimize the QUBO objective combining singleton utility, pairwise redundancy penalty, and a cardinality penalty that enforces exactly K selected sites.

100 candidate catalyst sites; select K=25 with beta=5 pairwise coupling (hard synthetic regime).

Exactly-K cardinality: select precisely 25 candidates.
No additional explicit constraints; redundancy is encoded as a pairwise QUBO penalty.

Approach

CGA GPU annealer

Quantum-Inspired

A third-party GPU-accelerated simulated annealing service used as an independent annealing reference to compare against classical greedy, SA, and D-Wave Hybrid on the identical QUBO instance.

timeout_sec60

server_urihttp://60.250.149.247:8080

protocolSocket.IO

lambda_K5

beta5

K25

Alternative	Why not
Greedy marginal selection	Faster and feasible, but yields a higher objective (5.3467) than CGA (4.7700) in this hard regime.
D-Wave Hybrid	Achieves the certified optimum vicinity (0.5074) but requires D-Wave Leap access; CGA was evaluated as an additional independent annealing reference.
Gurobi MIQP	Certified optimum (0.5074) used as the bound/reference; not a heuristic solver being compared.

Figures

Solver gap to best-known

Bar chart showing solver gaps across N=50,100,200,272 for greedy, SA, SQA, Fujitsu DA, HiGHS, D-Wave Hybrid, D-Wave QPU, Gurobi, and CGA GPU.

Source data: reports/phase8_solvers_full_canonical.csv

Bar chart of solver gap to best-known energy on identical locked QUBO instances (K=10, beta=1), including CGA GPU annealer.

Reliability

Runs: 1
Median: 4.770031153763284
Range: 4.770031153763284 – 4.770031153763284

Optimality

Bound: 0.5073599142524472
Achieved: 4.770031153763284
Gap: 840.1671318065263%

CGA returns a feasible solution but leaves a large gap to the certified optimum in this strongly coupled, high-selection-fraction regime.

Internal Notes

What worked: The Socket.IO submission and decode pipeline successfully returned feasible selections on all tested locked and hard cells.

What didn't: CGA did not close the gap to Gurobi or D-Wave Hybrid in the hard regime; its objective was comparable to or worse than greedy except on hard_N100K25B5 where it modestly beat greedy.

Next time: Run multiple seeds/repeats per cell, sweep anneal budgets (30/60/120 s), and compare CGA against Fujitsu DA and D-Wave Hybrid under identical time limits.

Reproduce

Commit: unknown
Command: cd /research/projects/catalyst-hpc-qml && python experiments/26_cga_bench.py