Embedded Files
Molecular Machine Learning
Molecular Machine Learning
We develop AI that learns from molecular and materials data to predict properties, guide simulations, and design promising new compounds. Our research connects five complementary directions:
Property prediction
Learn useful representations from unlabeled data to predict molecular and materials properties with fewer labeled examples.
Molecular simulations
Build neural potentials and scalable computational tools to accelerate atomistic modeling.
Optimization of molecules and materials
Use reinforcement learning and AI agents to explore candidates and refine designs through computational feedback.
De novo generation
Develop generative models, including MOFGPT, that propose new structures guided by target properties.
Interpretability and scientific insight
Combine domain knowledge with interpretable models to understand the relationships between structure and function.
Our research at MAIL
Our research at MAIL
From learning to targeted generation. MOFGPT combines pretraining and reinforcement learning (top). Property distributions below illustrate targeted gas adsorption and band-gap generation. Paper · Figs. 1 & 3
Chemical Language Model
Chemical Language Model
Chemical language models connect chemical representations with material properties. TransPolymer learns polymer representations for property prediction, while MOFormer learns representations of metal–organic frameworks (MOFs).
MOFGPT: from property prediction to generative design
MOFGPT combines a transformer trained on MOFid sequences, MOFormer property predictions, and reinforcement-learning feedback to propose MOFs with target properties. This connects representation learning to inverse design: generating candidates for further computational screening and experimental validation.
Self-Supervised Learning
Self-Supervised Learning
Self-supervised learning (SSL) addresses the challenge of limited labeled data in chemistry and materials science, allowing for improved accuracy and generalization of ML models. We develop various SSL models for molecule/material science.
AI in Material Design
AI in Material Design
We develop deep reinforcement learning (DRL) agents to learn from interactions with the environment, such as simulating material structures and properties, and iteratively optimize material designs based on simulation feedback:
Adsorb-Agent: autonomous search for stable adsorption configurations
Adsorb-Agent uses a large language model to propose promising molecule–surface configurations for computational evaluation. By focusing the search on informative candidates, it reduces the number of initial configurations needed to explore adsorption energies, connecting AI reasoning with catalyst discovery.
Predicting energies and generating structures. Below: graph-assisted CatBERTa learning, followed by CrystaLLM-generated adsorption structures alongside reference configurations. Ock et al., 2024 · Figs. 1 & 4
Property prediction
Property prediction
We introduce domain knowledge and interpretable mechanisms to “black-box” neural networks for the purpose of building more accurate models and providing insights to researchers:
Molecular Simulation
Molecular Simulation
We develop AI methods that accelerate molecular simulations and use molecular dynamics to study transport, membranes, and nanoscale sensing.
DistMLIP: scaling machine learning interatomic potentials
DistMLIP distributes interatomic-potential inference across multiple GPUs through graph partitioning. Its flexible interface supports existing models, including CHGNet, MACE, TensorNet, and eSEN, extending the size of atomistic systems that can be evaluated with machine learning potentials.
To know more about AI and Material Discovery, you can watch these videos
Google Sites
Report abuse