GraphBU: Generation of MILP instances with native graph blocks

Generating mixed-integer linear programming (MILP) instances is a key challenge in solver development and artificial intelligence model training. Real-world problems often come from private or highly specific applications, limiting the availability of training data. Until now, existing generators relied on templates, summary statistics, or local edits to the problem graph, but none captured how a subpart connects to the rest of the instance. GraphBU proposes a native graph approach where the basic unit is a local subproblem plus its coupling interface. This allows promoting nodes to master constraints or boundary variables, and then swapping compatible blocks, preserving the structure that solvers and learned policies need. Results show a statistical similarity of 93.4%, a feasibility of 96.7%, and an 8% improvement in Predict-and-Search training. This breakthrough opens the door to tailored applications in optimization, where the quality of synthetic data is critical. At Q2BSTUDIO, we understand the importance of building robust systems that integrate artificial intelligence with realistic data. That is why we offer AI services for businesses that combine advanced models with synthetic scenario generation. Additionally, our custom software solutions enable the implementation of proprietary generators like GraphBU in production environments, whether on AWS and Azure cloud services or with cybersecurity layers to protect sensitive data. The ability to create viable and representative MILP instances also enhances the use of AI agents in planning and logistics, and integrates with business intelligence tools such as Power BI to visualize the impact of decisions. Ultimately, GraphBU exemplifies how innovation in data structures can transform industrial optimization, and at Q2BSTUDIO we are ready to help companies capitalize on these technologies.