Large Language Models and Machine Learning Framework for Predicting Dental Ceramics Performance.

Area of Science:

• Materials Science
• Biomaterials Engineering
• Artificial Intelligence in Materials Science

Background:

• Dental all-ceramic restorations frequently fail due to clinical fractures.
• Improved mechanical properties and durability of dental ceramics are crucial for restoration longevity.
• Current material design and discovery processes are time-consuming and labor-intensive.

Purpose of the Study:

• To develop a large language model (LLM)-based framework for automated construction of a structured dental ceramic database.
• To integrate this database with machine learning (ML) for predicting material properties.
• To accelerate the design and optimization of novel dental ceramic materials.

Main Methods:

• Employed LLMs (Llama, Qwen, DeepSeek) for literature mining, including text and tabular data extraction.
• Developed an automated pipeline to systematically extract and structure compositional and performance data from research articles.
• Trained ten ML algorithms on the curated database to create predictive models for ceramic performance.

Main Results:

• LLM-based text classification achieved high accuracy (F1 score > 0.89).
• Machine learning models accurately predicted flexural strength (best Extra Trees model F1 = 0.928).
• SHAP analysis identified key compositional contributors (ZrO₂, SiO₂), guiding material optimization.

Conclusions:

• An AI-driven pipeline combining LLM data extraction and ML modeling offers a scalable approach for dental material discovery.
• This methodology accelerates the identification and optimization of dental ceramics and other biomaterials.
• Advanced LLMs and ML models hold significant potential for restorative dentistry and materials research.