Amer El-Samman
Physics-Backed ML Models in Chemistry
Compressing First-Principles Physics with Machine Learning
From 118 elements, quantum mechanics provides the rules that shape our universe. These rules determine how we treat disease, store energy, design materials, and process information. But first principles physics is enormous in scale. Even a modest molecule like caffeine requires an exact quantum calculation so large that no existing computer could store the full solution. Modern machine learning, trained on quantum mechanical data, offers a practical way through this. It compresses memory heavy quantum information into high quality statistical representations that remain faithful to structure, reflect the language of reactions, and can be transferred to new tasks with only small datasets.
The Hidden Language of AI in Chemistry: Unifying Graph and Language Models of Chemistry
The decision making frameworks of graph neural network models contain a hidden map of reaction formulae. Much like how natural language models capture relationships between words through vector arithmetic, where “King” minus “Man” plus “Woman” yields “Queen,” graph based models appear to learn an analogous language for chemistry. In this representation space, the model encodes reactions such as “Hydrogen” plus “Oxygen” resulting in “Water,” revealing that fundamental chemical transformations emerge naturally within the geometry of the learned graph embeddings.
Understanding the Internal Model Learned by Graph Neural Networks on Quantum Chemistry
Graph neural networks are among the most sophisticated probabilistic models used in chemistry, yet the principles guiding their internal decision making often remain obscure. Understanding their learned representations is essential for revealing the hidden logic within these complex GNN models. This work explores how to access and interpret GNN’s deeply embedded decision making frameworks.
Highlighted Projects
How Does Message-Passing Work in Graph Neural Networks Trained on Quantum Data?
How does message passing in graph neural networks constructs a chemical representation from local atomic environments? By learning how each atom’s contribution propagates through the molecular graph, we learn how graph model develop a latent space that captures electronic structure, functional groups, and reaction relevant patterns. This structure explains why these models can transfer learning to new tasks, enabling accurate prediction of diverse chemical properties and even recovery of reaction formulas.
Transfer Learning From Latent Space of Graph Neural Networks Trained on Quantum Data
Does a graph neural network trained on quantum chemical data compress a representation that extends beyond its original task. By probing transfer to properties such as electron occupancy, NMR, pKa, and solubility, we assess whether its compressed latent space can serve as a general purpose engine for chemical prediction and design.
