Multimodal Polymer Property Prediction

Overview

Using the Lumee-7B and Lumee-MM (multimodal) model variants, we developed a polymer property prediction system, a fusion architecture integrating natural language and molecular structure. This project demonstrates how domain-specific AI can be rapidly deployed using Lumee’s foundation models — even on limited data — to outperform traditional machine learning methods.

🔧 Method

We used:

• Textual embeddings from Lumee-7B to encode polymer SMILES/PSMILES representations

• 3D structural embeddings from cheminformatics models (e.g., Uni-Mol) integrated via Lumee-MM

• LoRA tuning to adapt models to 29k+ experimental and DFT-labeled data points for 22 polymer properties

The result was a multimodal fusion pipeline that supported rich reasoning from both language and structure — ideal for property regression and scientific discovery.

🧠 Key Properties Predicted

• Glass Transition (Tg), Melting Point (Tm), Band Gap (Egc/Egb)

• Mechanical: Young’s Modulus, Strength, Elongation

• Chemical: Density, Refractive Index, Conductivity

• Gas Permeabilities: CO₂, O₂, N₂, CH₄, He, H₂

📊 Performance Highlights

Compared to classical models, Lumee-based fusion models matched or exceeded SOTA benchmarks on most properties, even without pretraining on massive polymer datasets.

🔍 Why Lumee Worked

• LLM chemical knowledge embedded during pretraining on scientific corpora

• Long-context support (128k tokens) enabled full-molecule and sequence reasoning

• Modular fusion-ready design via Lumee-MM accelerated experimentation

• Token-level interpretability supported explainable AI insights

🚀 Applications

• AI-assisted polymer discovery

• Predictive simulation tools for materials R&D

• Educational or institutional research assistants

• Industry adoption in sustainable materials, electronics, coatings, and biomedical sectors

📩 Interested in applying Lumee to your research or chemistry pipelines?

→ hello@lumees.io

Scientific paper will be available soon.