Research Pillars

AI for Science

Despite significant advancements in scientific research and development in the past centuries, the exploration of new scientific frontiers is still a time-consuming and resource-intensive undertaking. The typical process of scientific discovery often involves multiple iterations of empirical data collection and rigorous analysis, taking years or even decades to unveil meaningful insights. Moreover, the escalating volume and complexity of data produced by modern experimental means presents further challenges for researchers to discern patterns and extract insightful conclusions from their findings. These challenges suggest the need for new tools and approaches to accelerate the scientific process.

AI for Science represents a transformative leap in how we approach and solve complex scientific challenges. AI-driven understanding and modelling of natural phenomena across multiple discipline could enable faster, more efficient, and more intricate scientific discoveries, thereby revolutionising scientific research into areas like drug and materials design, protein 3D structure prediction and knowledge discovery. 

AI for ScienceFig 1. AI for Science can be applied to scenarios of multiple scales ranging from microscopic level (e.g. small chemical molecules and larger macromolecules like proteins) to larger-scale systems (e.g. urban environment and human mobility), across both physical and biomedical sciences.

Research Focus

While AI holds immense potential, multiple challenges remain. 

Challenges of AIFig 2. Challenges of AI include limited availability of data (potentially due to high annotation costs), complex, non-linear relationships, prevalence of multimodal data, and need for consistency with natural laws.

To address the above limitations, we focus on the development of generalisable generative AI models and optimisation algorithms with the following key characteristics:

  • Interpretability and controllability in generation (including with human preferences)
  • Consistency with domain expertise, including adherence to physical laws (e.g. governing partial differential equations) and physical constraints
  • Cross-modal capabilities in models to maximise use of different kinds of available data

Applications

Materials Discovery

Conventional approaches for design and discovery rely on human intuition and experimental trial-and-error and are consequently expensive and slow. Our research can expedite discovery cycles in materials science and guide discovery of novel materials, including advanced materials such as lightweight high-entropy alloys (HEAs) and high-conductivity composites which are key to many industries. Specifically, we have developed interpretable, generalisable generative AI models and optimisation algorithms that integrate domain knowledge and physical constraints. These models can help to discover new materials with desired properties that can be validated in the lab. 

Accelerated Material Discovery with Generative AI modelsFig 3. Accelerated Material Discovery with Generative AI models.

Biomedical Sciences

With the development of models such as 1) cross-modality RNA foundation models to enhance the accuracy and relevance of RNA structure predictions; and 2) graph-based neural networks to learn complex relationships among data and identify biomedical interactions with high group cohesiveness, our AI techniques can push the boundaries of biomedical research and personalised medicine. In addition, we can also construct AI models to better model physical systems via physically-constrained models and resource-efficient cross-modal models.

Fig 4. RNA 3D structure prediction.