Skip to content

Latest commit

 

History

History
111 lines (70 loc) · 12.1 KB

File metadata and controls

111 lines (70 loc) · 12.1 KB

LLMater 🧪 Data Science in a Molecular Universe

DOI

Supervisor: Chao Zhang (Department of Chemistry, Uppsala University)

Group Members: André Ramos Ekengren, Ramkishor Prabhu Ramlal, Yue Zhou


Table of contents


Topic Overview

Large language models (LLMs) have shown impressive capabilities in domains such as natural language processing, code generation, and reasoning. In parallel, chemistry and materials science are becoming increasingly data-driven, with growing databases of reactions, molecules, and material properties. Bridging these two developments creates an opportunity: if molecular information can be expressed in a form that is compatible with LLMs, then models originally designed for human language might also support molecular property prediction and materials design.

In this project we focus on molecular prediction problems in material design that are relevant for battery research. Electrolytes play a central role in battery performance, safety, and lifetime, yet their design space is extremely large. Traditional machine-learning models have been successfully used for some property-prediction tasks, but they often require carefully engineered descriptors and do not easily generalize across different tasks or datasets. LLMs offer an alternative in which the same transformer architecture can be applied to a wide range of problems once the molecules are encoded in a suitable way.


Literature Survey (from 2022)

As a first step, we conducted a literature survey on LLM applications in chemistry and materials science (primarily from 2022 onwards). The survey covers:

  • Property prediction from molecular or materials representations, often using SMILES or graph-based encodings.
  • Generative design of molecules and materials with targeted properties.
  • Studies that probe the limitations of LLMs in scientific and multimodal settings.

The relevant papers are collected and summarized in a the table below.

Model name Architecture Data link Description Adaptation Purpose Year DOI Github link
LLM-prop T5 (encoder) Benchmark data predicting the properties of crystalline materials using large language models Fine-tuning Regression 2025 DOI Github
N/A GPT-J-6B, Llama-3.1-8B, Mistral-7B Github Assessment of fine-tuned large language models for real-world chemistry and material science applications Fine-tuning Classification Regression Inverse design 2025 DOI Github
MolRAG Llama-3-8B-Instruct GPT-4o Qwen2.5-7B-Instruct N/A N/A N/A N/A N/A DOI Github, Code not available yet.
ElaTBot ElaTBot-DFT Llama2-7b, GPT-4o Data Large language models for material property predictions: elastic constant tensor prediction and materials design Fine-tuning, (LORA+), Prompt engineering, RAG Regression, Inverse design / Material generation 2025 DOI Github
N/A LLaMA 3 N/A Regression with Large Language Models for Materials and Molecular Property Prediction Fine-tuning Regression 2024 DOI Code and data: part 1 Code and data: part 2
GPT 4 for Chemistry GPT 4 N/A ExploringGPT-4’s potentialin chemical tasks, such as foundational chemistry knowledge, cheminformatics, data analysis,problem prediction, and proposal abilities Prompt engineering N/A 2023 DOI No github
N/A GPT 3.5 No clear data source Polymer Solubility Prediction Using Large Language Models Fine-tuning Classification 2025 DOI No github
N/A Gemini 1.5 Github Can large language models predict the hydrophobicity of metal–organic frameworks? Fine-tuning Classification 2025 DOI Github
LLM4Mat-bench LLM-Prop, MatBERT, Llama 2-7b-chat, Data and checkpoints Benchmarking large language models for materials property prediction Fine-tuning, In-context learning Classification Regression 2025 DOI Github
PolyLLMem Llama-3 Uni-Mol Github Multimodal Machine Learning with Large Language Embedding Model for Polymer Property Prediction Fine-tuning, LoRA Regression 2025 DOI Github
gptchem GPT3 The repository includes example scripts and demo data sources Natural-language pretrained Prompt engineering, not for LoRA Classification Regression 2022 DOI Github
ChemLLM InternLM-2 ChemData,ChemBench Pretrained for chemistry Fine-tuning,open weights(LoRA applicable) Classification 2024 DOI Github
chemlift General LLMs “The framework can be used with any HuggingFace model that accepts text input.” N/A Natural-language pretrained Prompt engineering(Prompt + RAG) Classification Regression 2023 DOI Github
De novo Design of Polymer Electrolytes with High Conductivity using GPT-based and Diffusion-based Generative Models minGPT HTP-MD polymer-electrolyte dataset (SMILES) Chemistry-pretrained SMILES-based generative Fine-tuning hyperparameter grid search provided. (LoRA not described.) Inverse design 2024 DOI Github
ChemLoRA General LLMs”GPT-3 via API; any HuggingFace base such as GPT-2) with LoRA/PEFT adapters. QM9-G4MP2 small-molecule dataset(SMILES) Natural-language pretrained Prompt engineering (GPTChem-style prompts) **,**Fine-tuning(LoRA) Regression(molecular energy predictions) 2023 N/A Github
Text2Concrete GPT-3.5-turbo Data Known,Relevant,Random ( PubChem molecules) Natural-language pretrained(Gaussian Process Regression and Random Forest for benchmarks) Prompt engineering Regression 2023 DOI Github
Molecule-Discovery-by-Context ScholarBERT literature corpus(for context) and Known,Relevant,Random ( PubChem molecules) Natural-language pretrained Prompt engineering(prompt, RAG) Ranking for candidate discovery 2023 DOI Github
Using GPT-4 in parameter selection of polymer informatics: improving predictive accuracy amidst data scarcity and ‘Ugly Duckling’ dilemma GPT-4 Data GPT4 is used to predict the refractive index of a molecule and select necessary descriptors. Comparative study N/A Prediction Descriptor 2023 DOI Github
SMI-TED-IC SMI-TED Github Chemical foundation model-guided design of high ionic conductivity electrolyte formulations Fine-tuning Regression 2025 DOI Github
Leveraging GPT-4 to transform chemistry from paper to practice GPT-4 Data 1. Using GPT 4 to read scientific literature and generate actionable steps 2.Generate a script to communicate the steps to EasyMax Reactor N/A N/A 2024 N/A Github
ChemLLMBench GPT-4, GPT-3, Davinci-003, Llama, Galactica Data What can Large Language Models do in chemistry? A comprehensive benchmark on eight tasks In-context learning Generation, Classification, Ranking 2023 DOI Github

Methodologies

Within this context, our project addresses three main tasks:

  1. Molecular property prediction
    Use LLMs to predict molecular properties, for example the conductivity of candidate electrolyte molecules.

  2. Classification
    Formulate classification problems such as distinguishing high-conductivity from low-conductivity molecules or grouping molecules by other experimentally relevant labels.

  3. Molecule generation
    Explore generative modelling, where the goal is to generate new SMILES strings that are chemically valid and have desirable predicted properties. These tasks together span both predictive and generative aspects of molecular design.


Benchmark Results

Each group member re-implemented selected models from the literature, covering property prediction, classification, and molecule generation tasks. These implementations serve both as baselines and as starting points for small methodological variations and improvements.

We then experimented with commercial LLMs (e.g., GPT-4) and selected a common open-source backbone, Llama-3.2-3B-Instruct, as our main model. Across the three subtasks, each group member implements and evaluates the model against their respective literature baseline, and explores modifications to improve performance.

Code for the three subtasks is organised as:

  • ./regression/ – LLM and baseline models for regression-style property prediction (implemented by André Ramos Ekengren)
  • ./classification/ – LLM and baseline models for molecular classification (implemented by Ramkishor Prabhu Ramlal)
  • ./inverse-design/ – LLM-based SMILES generation and post-processing (implemented by Yue Zhou)

Future Work

  • Investigating improved fine-tuning and evaluation strategies, including robustness in low-data regimes and better uncertainty estimates.
  • Developing more targeted generative workflows that couple LLM-based generation with domain-specific filters or chemistry-based checks.

Contributions and suggestions are welcome, especially regarding datasets, baselines, and evaluation protocols.


License

This project is released under the MIT License.