Entering the Fujitsu Challenge
We joined the Fujitsu Quantum Simulator Challenge with a practical question: how far can a carefully formulated QAOA workflow take us when applied to energy optimisation? Our aim is not to declare a breakthrough. It is to learn where hybrid optimisation is useful, where it becomes difficult, and what evidence is needed before any result can be trusted.
Starting locally with Qulacs
We began close to the metal, on a local workstation. Qulacs gave us a fast way to learn the full QAOA loop: translate an energy decision into a QUBO, build the corresponding cost model, explore circuit depth and parameters, sample candidate solutions, and interpret the output. Small problems were a feature, not a limitation. They let us inspect behaviour closely and compare each result with classical baselines before adding more complexity.
Moving towards parallel simulation
As the model grows, simulation costs rise quickly. That led us to Qulacs with MPI. Distributed simulation is not simply a larger version of a local run: state must be partitioned, processes must communicate, and memory, runtime and reproducibility all need fresh attention. We are learning to separate algorithmic improvements from gains — or losses — caused by configuration and communication overhead.
The Fujitsu FX700/A64FX stage
Moving from our local environment to the Fujitsu FX700 changed the shape of the work. The A64FX architecture asks us to revisit builds, libraries, process placement and resource use rather than assuming that local settings will transfer unchanged. This stage is about making the workflow portable and observable first. Only then can we benchmark it fairly.
Exploring QARP
QARP is another part of the Fujitsu environment we are now exploring. At this point, we are investigating how it can sit alongside QAOA, Qulacs and MPI in a coherent experimental workflow. We are deliberately keeping that exploration high-level while we establish repeatable inputs, outputs and validation checks. The role of each component should be demonstrated by evidence, not inferred from its label.
Challenges and lessons so far
The hardest questions often appear before a simulator starts. A useful QUBO must represent the energy objective without allowing constraint penalties to overwhelm it. Parameter optimisation can find a low energy value without guaranteeing a meaningful operational solution. Every candidate therefore needs feasibility checks and comparison with a classical method. On the infrastructure side, MPI configuration, process counts and communication patterns can change performance substantially. Fair comparison means controlling problem instances, seeds, stopping conditions and measurement boundaries — and recording unsuccessful runs as carefully as promising ones.
Current progress and next steps
We now have the stages of the workflow in view: local modelling, small QAOA experiments, classical reference solutions, MPI preparation and the transition to the Fujitsu environment. The work is still in progress. Our next experiments will focus on reproducible execution across environments, validation of feasible solutions, controlled scaling tests and a clearer understanding of where QARP belongs in the pipeline. We will share conclusions only when the comparisons are consistent enough to support them.