(Source: wikipedia.com)

Home
AI4Science page


 

 AI4Science Projects

Goal-driven crystal structure generation

Designing crystals is difficult as it requires deep knowledge of materials science and exploration of a massive combinatorial space. This project develops AI-driven methods for rapid exploration of the crystal space towards targeted properties while satisfying domain-specific constraints. Our approach combines generative models with physics-informed neural networks to navigate the vast space of possible crystal structures, considering atomic composition, symmetry groups, and lattice parameters. The framework incorporates materials science principles like electronegativity rules, size effects, and thermodynamic stability to ensure generated structures are both novel and synthesizable. This enables efficient discovery of new crystals for applications ranging from semiconductors to energy storage materials.

The examples of DFT-verified stable structures (E_hull = 0 eV/atom).

Designing MOF for CO2 capturing

This project develops Generative AI to accelerate the discovery of Metal-Organic Frameworks (MOFs) optimized for CO2 capture. By combining deep generative models with high-throughput screening, we develop an AI framework that learns the complex relationships between MOF structure, chemical composition, and CO2 adsorption capabilities. The system generates novel MOF candidates by exploring the vast chemical space of metal nodes and organic linkers while ensuring synthetic feasibility and stability. Our approach incorporates domain knowledge about pore size distributions, surface area constraints, and binding site energetics to prioritize structures with high CO2 selectivity and working capacity under industrial conditions.

(a) Example MOF structure (b)(c) Example metal clusters discovered.

Chemical reaction prediction

Predicting chemical reactions is central to chemistry, impacting fields from drug synthesis to materials processing. In this project we reframe the reaction prediction problem as graph morphism, where a collection of reactants and catalysts form a supra-graph with temporarily disjoint sub-graphs. Reactions are viewed as a sequential decision problem, where each bond change is an action in the reaction pathway. Using reinforcement learning and graph neural networks, our model learns to predict likely reaction mechanisms by sequentially modifying molecular bonds, considering aspects such as electronic effects, steric hindrance, and thermodynamic feasibility. This approach enables prediction of complex reaction pathways, yields, and side products while providing mechanistic insights into chemical transformations.

Reaction represented as a set of graph transformations from reactants (leftmost) to
products (rightmost).

Prediction of chemical-chemical interaction

Chemicals are rarely used in isolation but are typically surrounded by other chemicals in solutions, mixtures, and reactions. This project builds flexible machine learning models that predict interactions between any subset of chemicals in complex environments. By leveraging graph neural networks, attention mechanisms and neural computers, our approach captures both pairwise and higher-order interactions between molecular species. The framework can accounts for any contexts like concentration effects, pH conditions, and environmental factors that influence chemical behavior. This enables accurate prediction of properties like solubility, reactivity, and stability in multi-component systems, crucial for applications in drug development, materials synthesis, and chemical process optimization.

Flexible multiple molecules interaction modelling, aiming at answering multiple queries about the molecular system. The underlying model is Relational Dynamic Memory Networks.

Foundation Models for science

The scientific enterprise has generated massive empirical data from simulations and experimental studies. Much of the knowledge is documented in the scientific literature, in the form of textual description, mathematical equations, diagrams and tables. All of these knowledge sources can be intergated into an associative memory to be retrieved later in the form of Foundation Models (FMs). Thi research program aims at leveraging recent advances in Large Language Models (LLMs) to build scientific FMs, which will later be used by AI Scientist agents in scientific workflows. Left: DALL·E 3 illustration of Foundation Models, compressing all scientific knowledge into its parameters.

Physics-informed GNNs for materials

This project develops physics-informed graph neural networks (PiGNNs) to model materials, incorporating fundamental physical laws such as symmetries and conservation directly into the neural architecture. Unlike traditional GNNs, physical priors are encoded through custom loss functions and specialized message-passing operations. This approach aims to improve prediction accuracy for properties like formation energy, bandgap, and elastic moduli while ensuring physically consistent results, even with limited training data.

Embedding material graphs using the electron-ion potential.

Modelling crystals plasticity

This project develops deep neural networks to model stress-strain relationships in polycrystalline materials, where multiple grain types coexist with distinct crystallographic orientations. Our approach captures the complex interplay between individual grain deformation mechanisms, grain boundary interactions, and overall mechanical response. By incorporating microstructural features like grain size distributions, misorientation angles, and texture evolution, the model predicts heterogeneous plastic deformation across different grain populations. The framework accounts for grain-specific slip systems, local strain incompatibilities at boundaries, and texture-dependent hardening behaviors to enable accurate prediction of polycrystalline materials' mechanical properties under various loading conditions.

Crystal microstructures.

Drug-protein binding prediction

This project aims at inventing new data-efficient AI models to precisely predict the location and strength of binding between drug molecules and target proteins, a critical challenge in drug discovery. By developing advanced neural architectures that capture both 2D/3D drug conformations and protein structures, we model the complex physicochemical interactions at binding sites. Our approach integrates multiple learning strategies: representation learning for 2D/3D molecular structures, attention mechanisms for binding site identification, transfer learning from related protein families, and interpretability methods to explain predictions. The framework incorporates biophysical constraints and leverages limited experimental data to achieve accurate binding affinity predictions while providing mechanistic insights into drug-protein interactions.

Attention values at predicted binding sites of MST1 target.

Deep hybrid generative-discriminative model for inverse design

This project aims at developing general data-driven techniques for predicting the design parameters for any target in a single step. This poses two technical challenges: the first caused due to one-to-many mapping when learning the inverse problem and the second caused due to an user specifying the target specifications only partially. To overcome the challenges, we formulate this problem as conditional density estimation under high-dimensional setting with incomplete input and multimodal output.

CVAE-MDN, where x: input design, v: specified target component, h: unspecified part, and z: latent variable. Left: Search speed comparison of the proposed CVAE-MDN with other search techniques.

Multi-target molecular property prediction

Molecules have multiple properties of interest, and often all of them must be satisfactory for practical use in applications like drug discovery and materials design. This project develops new scalable and explainable techniques based on Graph Neural Networks to predict hundreds of properties simultaneously, leveraging the strong correlations between them. By encoding molecular structures as graphs and utilizing advanced attention mechanisms, our models capture complex structure-property relationships across diverse chemical spaces. The framework achieves state-of-the-art accuracy while providing interpretable insights into property predictions. Notably, the models are capable of zero-shot learning, enabling prediction of new properties using only their textual descriptions, thus reducing the need for extensive experimental data.

Substructures of a molecule corresponding to 8 prediction targets (PubChem SID: 491286)