Your Tasks, Core Themes

Candidates will have the opportunity to pursue one or more of the following core themes:

1.     Large Pre-Trained Models for Multi-Modal Reasoning

This project explores the integration of graph-based foundation models (e.g., knowledge graphs) with large language models (LLMs) to build AI systems capable of reasoning across diverse scientific data modalities such as structured databases, ontologies, and unstructured text. The goal is to enable automated hypothesis generation and obtain scientific insights, with potential applications in biomedical discovery, literature mining, and cross-domain reasoning.

2.     Deep Surrogate Models for Complex Scientific Systems

This theme focuses on building deep learning-based surrogate models that approximate complex physical and biological systems traditionally modeled by PDEs or other computationally expensive simulations. By incorporating physical priors such as conservation laws and symmetries into architectures like neural operators, physics-informed neural networks (PINNs), and graph-based solvers, the project aims to accelerate simulations in areas including protein folding, gene regulation, drug–target interactions, and multi-scale biomedical modeling.

3.     Deep Learning on Graphs with Higher-Order Structures

This project develops neural architectures capable of modeling higher-order relational structures such as hypergraphs and simplicial complexes, which are prevalent in complex biological systems like protein assemblies, signaling pathways, and metabolic networks. The goal is to build foundation models capable of learning from richly structured or semi-structured data where traditional graph neural networks may fall short, enabling better representation, inference, and discovery in complex domains.

4.     Theoretical Foundations of Geometric Deep Learning

This theme addresses foundational questions in geometric deep learning, including the expressiveness of graph neural networks under symmetry constraints, their optimization dynamics, and their generalization behavior—particularly in low-data or out-of-distribution settings. The work combines formal theoretical analysis with practically motivated case studies, offering a strong foundation for researchers interested in advancing the mathematical understanding of geometric deep learning.