Using large language models for enhancing accessibility for Monte Carlo photon transport simulations and beyond

Area of Science:

• Biomedical optics
• Computational biophysics
• Scientific software development

Background:

• Computational modeling and simulation software are crucial for biomedical optics research.
• Novice users face challenges understanding complex physical problems and software configurations.
• Large language models (LLMs) offer potential for natural language interaction but struggle with output reproducibility in technical applications.

Purpose of the Study:

• To investigate the use of LLMs in quantitative biophotonics simulation tools.
• To enable novice users to build complex photon simulations via natural language descriptions.
• To bridge the gap between natural language and advanced simulation software.

Main Methods:

• Explored prompt engineering strategies to constrain LLM outputs using a data schema and modular architecture.
• Implemented deterministic validation to ensure correctness and reproducibility of LLM-generated simulation inputs.
• Developed an LLM interface, MCX-LLM, for the Monte Carlo eXtreme (MCX) photon transport simulator.

Main Results:

• MCX-LLM achieved 98% accuracy and 99% repeatability across 33 diverse natural language simulation descriptions.
• The framework demonstrated 100% success on 20 unconstrained real-world prompts, handling various linguistic styles.
• The LLM interface showed generality by producing valid inputs for a finite-element-based diffusion solver.

Conclusions:

• Combining LLMs with structured constraints makes complex scientific tools more accessible.
• The developed framework ensures reliability and technical correctness for rigorous scientific research.
• MCX-LLM is integrated with MCX Cloud, available at https://mcx.space/cloud.