I’m passionate about advancing materials science by integrating state of the art AI techniques. My contributions span from materials representation learning to NLP. 🌟

LLMForge: Language Models Enable Data-Augmented Synthesis Planning for Inorganic Materials

• Benchmarking of seven state-of-the-art LMs on precursor recommendation and synthesis-condition regression, achieving up to 53.8% Top-1 precursor accuracy and MAEs below 126 °C. Generated 28,548 synthetic solid-state recipes via OpenAI GPT-4.1, expanding current dataset size by 616 % over literature-mined protocols.
• Methods: Large Language Models, Data Augmentation, Transformer Pretraining, Ensemble Modeling
• Repository: Project GitHub & Dataset

MTENCODER: A Multi-task Pretrained Transformer Encoder for Materials Representation Learning

• Developed a transformer-based encoder co-trained across diverse materials properties and a denoising objective, resulting in robust and generalizable materials representations.
• Methods: Multi-task Learning, Transformer Architecture, Denoising Autoencoders
• Paper: (NeurIPS AI4Mat) MTENCODER: A Multi-task Pretrained Transformer Encoder for Materials Representation Learning
  

Reaction Graph Networks for Inorganic Synthesis Condition Prediction

• Combined a Deep Learned Material Encoder with Graph Networks to model precursor interactions in inorganic reactions, enabling the prediction of synthesis conditions for solid-state materials.
• Methods: Graph Neural Networks, Inorganic Reaction Modeling, Synthesis Condition Prediction
• Paper: (NeurIPS AI4Mat) Reaction Graph Networks for Inorganic Synthesis Condition Prediction

A Chemically-Guided Generative Diffusion Model for Materials Synthesis Planning

• Introduces a generative model utilizing diffusion processes to predict viable synthesis routes for zeolite materials, considering the complex one-to-many relationships between structure and synthesis.
• Methods: Diffusion Models, Generative Modeling, Materials Synthesis Prediction
• Paper: (NeurIPS AI4Mat Spotlight) A Chemically-Guided Generative Diffusion Model for Materials Synthesis Planning

Augmenting Scientific Creativity with Retrieval across Knowledge Domains

• Developed an exploratory search system enabling scientists to select core text from a paper abstract and retrieve cross-domain papers with high similarity, facilitating knowledge transfer across scientific domains.
• Methods: Sentence Transformers, Clustering Techniques & Metrics
• Paper: (NAACL 2022) Augmenting Scientific Creativity with Retrieval across Knowledge Domains
• Code: GitHub Repository

Extracting a Database of Challenges and Mitigation Strategies for Sodium-ion Battery Development

• Created a detailed database highlighting performance and synthesis challenges in sodium-ion battery (SIB) cathode materials, alongside proposed mitigation strategies, to accelerate SIB research and development.
• Methods: Open Information Extraction, Named Entity Recognition, Coreference Resolution
• Paper: (NeurIPS AI4Mat) Extracting a Database of Challenges and Mitigation Strategies for Sodium-ion Battery Development
• Code: GitHub Repository
  

Regress, Don't Guess – A Regression-like Loss on Number Tokens for Language Models

• Introduced a novel loss function that enhance language models' numerical reasoning by considering the proximity between number tokens, thereby improving arithmetic capabilities.
• Methods: Autoregressive Transformers (T5)
• Paper: (NeurIPS MathAI) Regress, Don't Guess – A Regression-like Loss on Number Tokens for Language Models
• Code: GitHub Repository