Automatic User Interface Generation for Scientific Command-Line Applications Using Large Language Models

Abstract

Scientific software is available as command-line software with powerful computational power, but cannot be used by non-programming users. Although recent studies show the remarkable potential of the Large Language Models (LLMs) applied to the software development process, the potential of these models for modernizing the legacy scientific software in a systematic way has not been fully explored. Studies on the software engineering lifecycle are mostly limited to the generation of single code programs or prototype web applications. This paper introduces an AI-assisted software engineering framework to create graphical user interfaces (GUIs) for scientific command-line applications using several AI-centered software development tools. The proposed framework uses GPT-5.5, GITHUB COPILOT, GAALO, and Vercel v0 throughout the software development lifecycle to capture the requirements, generate user stories, design software interfaces, generate source code, debug, validate, and refine software. This methodology is based on Action Research, the LLM-generated objects are continually reviewed, integrated, and refined over five development cycles, and software correctness, architectural consistency is achieved. The framework is assessed by developing an open source Python library implementing classical heuristics for the Capacitated Vehicle Routing Problem (CVRP), “VeRyPy”. The generated desktop GUI can be used to configure the problem, select a heuristic algorithm, visualize the problem, report on the performance of the behavior, and export the results of the problem without altering the algorithms used for optimization. The case study illustrates how LLM can be used to support a large part of the construction of GUI, and conversely, how humans are still needed to perform the final GUI validation, integration, debugging and quality assurance. The results illustrate the power of a modern collaborative Human-AI workflow to update legacy scientific software, along with realworld problems with context-window size, validation, and the architectural differences of AI systems.