Results of the FY2025 MITOU Target Program: Software Development Utilizing Quantum Computing Technology
Release Date:Sep 30, 2026
In-page Links
Results of the FY2025 MITOU Target Program: Software Development Utilizing Quantum Computing Technology
Category 1: Software Development for Annealing Machines
PM: TANAKA Shu
Optimization of Distributed Renewable Energy Systems through Probabilistic Quantum Annealing Using Bayesian Factorization Machines
To rapidly optimize power supply and demand planning while accounting for uncertainty in renewable energy generation and demand forecasts, we developed BFMQA, a method that combines Bayesian Factorization Machines with quantum annealing. By extending conventional FMQA in a Bayesian manner and generating multiple QUBOs from the posterior distribution for parallel exploration, BFMQA achieves both diversity in search and convergence while suppressing stagnation in local optima. For power supply and demand planning, instead of directly optimizing a large number of computationally expensive scenarios, we implemented an iterative workflow in which the model is trained on a small number of scenarios and candidate solutions are then evaluated on a large number of scenarios. On benchmark functions, BFMQA showed more stable convergence than FMQA on highly multimodal problems, and performance improvements were also confirmed as the number of QUBOs increased. In evaluations of supply and demand planning under varying weather conditions and generation uncertainties, we also demonstrated that the expected cost could be brought progressively closer to the reference solution obtained using many scenarios.
Creators: ITO Yuya, KUROIWA Taihei
Exploring the Use of Quantum Annealing in Entertainment and the Potential of Games as a Means of Deploying Advanced Technologies in Society
To enable the general public to gain an intuitive understanding of quantum annealing through hands-on experience, we developed “AFTER・IMAGE,” a shooting game centered on evasion optimization, with an annealing engine serving as its game AI. In the game, an enemy robot uses annealing to optimize its posture and evade the player’s bullets. The system integrates a Unity-based frontend with a Python-based backend and uses Fixstars Amplify AE. To handle the enormous search space, we designed a three-stage model consisting of yaw-angle, axis, and limb models, represented joint angles as discrete states to formulate the problem as a QUBO, with the number of bullet hits as the sole objective to be minimized. Parallel processing and Numba JIT compilation enabled solve times of under five seconds, and about 200 people played the game at the Taipei Game Show. By visualizing the annealing results as the robot’s movements, the project demonstrated a pathway for helping people understand advanced technologies through entertainment.
Creator: KURAMOTO Kou, KANAI Keita, OOKAWA Takuto, HIRAI Nobuyuki
Development of Quantum-Annealing-Based Software for Perovskite Crystal Structure Exploration
I developed software for perovskite crystal structure exploration that reformulates the problem as a discrete combinatorial optimization task—selecting the arrangements and orientations of atoms and organic cations—rather than as continuous optimization of atomic coordinates. The resulting problem is solved using quantum annealing. In this approach, each lattice site in a 2 × 2 × 2 supercell is treated as a discrete occupancy site, with elemental configurations and molecular orientations represented by binary variables and formulated as a Quadratic Unconstrained Binary Optimization (QUBO) problem. Interatomic interactions are evaluated using the MACE machine-learning interatomic potential, while constraints on composition ratios and one-state-per-site occupancy are incorporated into the optimization. The software was implemented in Python using Streamlit and Fixstars Amplify, providing an end-to-end browser-based workflow from inputting search conditions and exploring candidate structures to exporting the resulting structures in CIF format. Whereas exhaustive evaluation of candidate structures can take several hours, the proposed approach identifies structures within the lowest-energy 1–2% of candidates in approximately 10–20 seconds. By efficiently narrowing the search space while suppressing physically unreasonable structures, the software can reduce the computational cost of subsequent first-principles calculations and accelerate materials research.
Creator: FUKASAWA Ryo
PM: TANAHASHI Kotaro
Development of a PCB Design Support Tool Using Annealing Machines
We developed a tool that automates PCB routing—the most labor-intensive stage of PCB design—using an annealing machine. Routing is a large-scale combinatorial optimization problem: several hundred pairs of start and end points must be connected within a limited area without crossings or short circuits. Our tool first generates a large set of candidate paths on a coarse mesh, then uses an annealing machine to select a mutually non-conflicting combination of them, thereby rapidly determining the global route of each connection. We confirmed that this approach yields high-quality routing while keeping computation time in check. The tool is designed for integration with KiCad, an open-source EDA tool, and we plan to release it as a cloud service.
Creators: KATO Shunsuke, NAGAYAMA Kotaka, TOYAMA Kota
Development of a QUBO Formulation Method Optimized for Embedding into Quantum Annealers
To avoid densely connected QUBO formulations, which pose a major obstacle when solving problems with quantum annealers, I developed a formulation method that transforms equality and inequality N-hot constraints into sparsely connected forms. Instead of directly squaring long expressions, the method introduces auxiliary variables and decomposes the constraints, reducing the number of couplings from the conventional O(N²) to O(N log N) for general N-hot constraints and to O(N) for one-hot constraints. This makes it possible to suppress the average chain length and the number of required physical qubits during embedding, thereby improving the ability to solve large-scale and complex problems with high accuracy. I released the result as the open-source library “sparse-qubo,” which supports both D-Wave’s dimod and the Fixstars Amplify SDK. In addition to being designed for easy adoption without major changes to existing code, the library also includes a visualizer that allows users to inspect decomposition structures and variable relationships. Its effectiveness was confirmed through shift scheduling problems, where it consistently produced high-quality solutions.
Creators: SUDA Kohei
PM: TAMURA Ryo
Development of an Automated QUBO Construction System Using Machine Learning
To automate QUBO formulation, which is a major bottleneck when solving combinatorial optimization problems with annealing machines, we developed a method that uses a binary autoencoder to compress feasible solutions into a low-dimensional binary latent representation while preserving their essential features. We further approximate the objective function in this latent space using a factorization machine, express the resulting approximation as a QUBO, and perform iterative optimization by incorporating feedback from candidate solution evaluations. We released Quron, an open-source Python library that provides a unified framework for the entire process. In experiments on the traveling salesman problem, the method compressed a 64-bit representation to 14 bits and was less prone to becoming trapped in local optima than existing rule-based approaches or random assignment methods, demonstrating its effectiveness in terms of both the feasibility rate and the efficiency of reaching optimal solutions.
Creators: ABE Tetsuro, YAMASHITA Masashi
Proof of Concept for a Ride-Sharing Application Using Annealing Machines
Using annealing machines, we developed an application that sequentially performs global optimization of ride-sharing dispatch for both on-demand and pooled transportation services. The system integrates a map-based UI built with React, a backend using Nest.js and PostgreSQL, and an optimization service based on FastAPI, Celery, and Amplify AE. Through asynchronous communication via WebSocket and batch processing at five-second intervals, it solves dynamically generated dispatch requests in parallel. We also implemented a dynamic simulator that generates virtual taxis and ride requests on OpenStreetMap, enabling evaluations under conditions that reflect real-world operational environments. Compared with simple classical assignment methods, the system reduced total taxi travel distance by as much as 26.4% while keeping the increase in response time to only a few seconds, and it can also flexibly accommodate additional constraints such as user preference conditions.
Creators: IDE Shunta, FUKUHARA Hiroki
Category 2: Software Development for Gate-based Quantum Computers
PM: TOKUNAGA Yuki
Demonstration of Quantum Layer-Integrated AI for Imperfect-Information Games and the Establishment of a Quantum Version of Kaggle
We developed QuAic, an experimental platform that enables users to learn quantum-classical hybrid machine learning through hands-on experience and comparative evaluation. The platform uses the imperfect-information game “Geister” as its theme, allowing users to design models that combine quantum and classical layers through the GUI-based HNN Composer, train them on Google Colab, and evaluate their performance through automated matches and rating after submission. The target function was narrowed to a module that estimates the colors of an opponent’s pieces from the board state, while action selection was handled by classical deep reinforcement learning. Color estimation using QNN improved the win rate compared with random estimation, demonstrating the effectiveness of a division-of-labor design in which the quantum layer is specialized for estimation. Going forward, we aim to build up design guidelines for quantum machine learning through long-term Kaggle-style competitions and knowledge-sharing features.
Creator: SOTOKAWA Kisho, EBI Hana, KAMEI Suzukaze
Development of a Compiler for Fault-tolerant Measurement-Based Quantum Computing
We developed end-to-end software for photonic quantum computers that consistently converts logical quantum circuits into execution command sequences for fault-tolerant measurement-based quantum computing (MBQC) and evaluates logical error rates. The software implements a stepwise compilation pipeline from quantum circuits to lattice-surgery instructions and then to MBQC command sequences, and connects to Stim so that the resulting circuits can be evaluated with its efficient Clifford simulator. In addition, the software supports photonics backend simulator where it maps optical parameters such as loss, fusion fidelity, and squeezing level to effective Pauli noise and erasure errors, enabling comparisons between GKP-based and FBQC-based schemes as well as threshold exploration. By using third-party lattice-surgery compiler backend, for a 4-qubit adder circuit, it evaluated a total of about 1.04 million commands, a maximum of 1,252 simultaneous qubits, and a circuit depth of 1,438, showing that it can be used to estimate required hardware performance and support design optimization.
Creators: FUKUSHIMA Masato, WATANABE Yuki, SASAKI Daichi, INOUE Shinichi, OKAZAKI Koichi, ITSUI Naoki
PM: FUJII Keisuke
Development of a Cooperative Digital Game Applying Concepts of Quantum Technology for Users Unfamiliar with Quantum Technology
We developed a cooperative digital game that enables even beginners unfamiliar with quantum technology to gain an intuitive understanding of the engineering technologies that support quantum computers. Built with Unity 2022.3 LTS, the game supports web browser, Windows, and Mac versions, and its maintainability was improved through a data-driven design based on external JSON and CSV files together with a dedicated configuration tool. Players advance development in a turn-based manner by strengthening core technologies such as cooling, vacuum systems, and noise control as skills, while enhancing quantum-specific elements as abilities, all while making choices about fundraising, research, and securing personnel, using indicators such as qubit count and error rate. The game implements six quantum computing approaches—superconducting, ion trap, neutral atom, semiconductor, single-photon, and continuous-variable—and reproduces differences in difficulty and contribution among them. It also includes a glossary function and future support for cooperative features, aiming to deepen structural understanding of quantum technology and improve technological literacy.
Creators: TATSUTA Yukina, YAGUCHI Kotomi, KAWATOMI Yuika
EntanglePlan: Development of a Strategic Entanglement Purification Tool for Quantum Networks
We developed EntanglePlan, a software framework that estimates the time-varying quality of quantum entanglement in quantum networks in real time and dynamically determines purification strategies. Using Stim, the system simulates Bell-pair generation, noise application, purification circuits, and measurement, and it employs a three-layer loop structure in which the Runner layer executes simulations, the Distimation layer performs state estimation, and the Planner layer determines the next purification policy. The framework allows users to configure noise models, generation rates, and target fidelity, and both the estimation method and purification scheme can be replaced as needed. While fixed-round purification can result in either insufficient quality or excessive resource consumption, we confirmed that an adaptive approach, which adjusts the number of purification rounds according to the estimation results, can stably maintain the target fidelity while reducing resource usage. Its novelty lies not in purification itself, but in presenting the operational layer for monitoring and controlling entanglement quality as software.
Creator: YOKOMORI Hikaru, KOYAMA Marii
PM: YAMAMOTO Naoki
Verification and Optimization of the Effectiveness of Reservoir Computing with Non-Markovian Open Quantum Systems
To verify the effectiveness of quantum reservoir computing (QRC) that explicitly incorporates non-Markovianity, we developed a simulation framework based on spin–spin systems and spin–boson systems. We quantified the memory effects generated by interactions with the environment using the BLP measure, and evaluations with STM and NARMA showed that the spin-environment model outperformed the model without an environment on both tasks, while the bosonic-environment model outperformed it on the NARMA task. Furthermore, in ECG200 classification, we found that even sparse sampling without measuring all time steps caused only a small drop in performance, and in particular, improved classification accuracy was confirmed under the condition with a spin environment. We also prepared demonstration notebooks in Jupyter, showing the design potential of using non-Markovian dissipation and environmental degrees of freedom not as sources of error, but as computational resources.
Creator: SASAKI Daiki, KOGA Ryosuke
Development of a Large-Scale Combinatorial Optimization Method Using Quantum Signal Processing
To advance the practical implementation of Grover Adaptive Search with Quantum Signal Processing (QSP-GAS), this project develops an oracle circuit construction method suitable for large-scale combinatorial optimization, together with a high-speed simulator that enables unified comparisons among multiple approaches. The proposed oracle circuit construction decomposes the cost Hamiltonian into sub-Hamiltonians and uses anticommuting Pauli strings to avoid the direct implementation of controlled time evolution, thereby limiting increases in circuit depth and CNOT gate count. Evaluation on 500 random objective functions of degree up to four shows that the proposed method achieves lower average circuit depth and CNOT gate count than conventional methods over a range of 4 to 24 qubits. The simulator specifies simulation conditions in YAML and automates method selection, initial-state configuration, evaluation of convergence behavior and circuit size, visualization, and data storage. In addition, acceleration methods that exploit the mathematical structures of QD-GAS and QSP-GAS without explicit circuit generation are implemented. These methods preserve convergence performance equivalent to that of Qiskit-based circuit simulations while reducing execution time by factors of approximately 247 for QD-GAS and 106 for QSP-GAS.
Creator: FUJIWARA Shintaro
Change log
-
Sep 30, 2026
Released this page