Artificial intelligence orchestration for text-based ultrasonic simulation via self-review by multi-large language model agents

Abstract

Abstract Widely used ultrasonic simulation systems often rely on complex graphical user interfaces (GUIs) or scripting, resulting in substantial time investments and reduced accessibility for new users. In this study, we propose a novel text-based simulation control architecture, which leverages a large language model (LLM) and the ground artificial intelligence (AI) approach to streamline the control of ultrasonic simulation systems. By modularizing the functionalities of the SimNDT program into discrete functions and enabling natural language-based command interpretation, the proposed method reduces the average simulation configuration time by approximately 75%. To further mitigate task failures in scenario generation using the LLM, we introduce the ground AI approach, which employs self-review mechanisms and multi-agent collaboration to improve task completion rates. In particular, when vectorized output lengths deviate from the standard, we regenerate outputs using multiple LLM agents, reducing the scenario generation error rate from 23.89 to 1.48% and enhancing reliability significantly. These advancements underscore the potential of AI-driven methods in reducing operational costs and enhancing reliability in simulation frameworks. By integrating text-based control and Ground AI mechanisms, the proposed approach provides an efficient and scalable alternative to traditional GUI-based control methods, particularly in time-sensitive applications such as digital twin systems.