Hyungjun Kim, Dohun Kang, Jaejun Lee, Jiwon Sun, and Seunghan Lee

Elif Ertekin Group, University of Illinois at Urbana-Champaign
Chris Wolverton Group, Northwestern University
Implementation by Jaejun Lee

• Project Overview
• Installation & Setup
• Core Components
• Performance Results

This repository demonstrates the superior performance and efficiency of TabPFN (large language model) when dealing with unseen small datasets. We leverage Retrieval-Augmented Generation (RAG) using OpenAI and Materials Project databases to extract domain-specific knowledge, taking advantage of TabPFN's in-context learning capabilities.

• RAG-Enhanced Feature Selection: AI-driven domain knowledge extraction for optimal descriptor selection
• Intelligent Clustering: Handles TabPFN's 10k sample limitation through chemistry-aware data segmentation
• Real-time Adaptability: Dynamic feature updates based on specific materials properties
• Efficient Training: Significantly faster than traditional LLM fine-tuning approaches

• Superior Performance: TabPFN fine-tuning outperforms LLM-based fine-tuning on unseen small datasets
• Training Efficiency: Significantly more efficient learning process compared to traditional approaches
• Robust Feature Handling: Leverages TabPFN's in-context learning to handle:
  • Variable feature lengths without retraining
  • High proportions of missing (NaN) values
  • Direct deployment with high performance expectations

• Python 3.12.11
• CUDA-compatible GPU (recommended for TabPFN)
• Google Colab or Colab Pro/Pro+ (recommended due to GPU memory requirements)

pip install pymatgen==2025.6.14
pip install mp-api==0.45.8
pip install openai
pip install tabpfn

or you can just use

pip install -r requirements.txt

1. OpenAI API Key

   • Get from: https://platform.openai.com/api-keys
   • Edit rag.py line 8: API_KEY = "your-openai-key-here"

2. Materials Project API Key

   • Get from: https://next-gen.materialsproject.org/api
   • Set as environment variable: export MP_API_KEY="your-mp-key-here"

# Clone repository
git clone https://github.com/Ahri111/ZEROMAT.git
cd ZEROMAT
conda create -n yourenv python=3.12.11
conda activate yourenv

# Install dependencies
pip install -r requirements.txt

Purpose: Intelligent feature selection using domain expertise from GPT-4

Key Features:

• Physics-informed feature recommendations
• Interactive chat interface for query refinement
• Automatic parsing and validation of recommended features
• Integration with Materials Project feature mapping

Usage:

python rag.py

print("Choose mode:")
print("1. Interactive chat")
print("2. Quick test")

Example Interaction:
Your question: What are the most important features for predicting bandgap in perovskites?

AI Response:
**RECOMMENDED FEATURES:**
1. formation_energy_per_atom - Thermodynamic stability indicator
2. density - Structural compactness affects electronic properties
3. band_gap - Direct electronic property correlation
4. dielectric_total - Electronic screening effects
5. bulk_modulus - Mechanical stiffness relates to bonding

These Recommended features will be saved as a name of feature_rocommendations.txt

Purpose: Production-ready data augmentation using Materials Project database

Key Features:

• Batch processing for large datasets (configurable batch sizes)
• Automatic material ID validation and cleaning
• Robust error handling and retry mechanisms

Usage:

export 
python production_mp_fetcher.py
export MP_API_KEY="your-materials-project-key"

Write all the requirements for implementing  production_mp_fetcher.py as follows:

Purpose: To evaluate the performance of our proposed mathod compared to Bert + Finetuning. Note that it is better to implement this Script at Colab rather than your own comptuter

Quick Start

• RAG Integration Impact: 43% improvement in R² score (0.5788 → 0.8261)
• Training Speed: 30x faster than LLM-Prop fine-tuning (38 min → 78.32 s)
• Memory Efficiency: 2-4x lower GPU memory requirements
• Accuracy: Best-in-class R² score of 0.8261 with RAG enhancement
• Error Reduction: 44% lower MAE compared to LLM-Prop + TabPFN alone

LLM-Prop + Fine-tuning (Baseline):

• Longest training time (38 minutes)
• Poorest performance (R² = 0.3881, MAE = 0.7999)
• High computational cost

LLM-Prop + TabPFN:

• 37x speed improvement over fine-tuning
• Moderate performance gains (R² = 0.5788)
• Significant memory efficiency

LLM-Prop + TabPFN + RAG (Our Method):

• Best overall performance (R² = 0.8261, MAE = 0.3652)
• Minimal additional training time (+16 seconds)
• Superior domain knowledge integration

• Dataset Size: Successfully handles ~40k materials
• Cluster Coherence: Maintains chemical similarity within clusters
• Scalability: Automatic splitting for large datasets
• Coverage: 98%+ successful materials property prediction