Research Interests

There are plenty of AI/ML in science applications today, especially for chemistry and material sciences. What are the challenges in methodology?

Traditional methods:

In typical discovery, traditional trial-and-error methods are reliable but extremely time-consuming and expensive. While traditional methods often face significant theory-to-practice gaps, recent attempts using standard machine learning approaches still remain limited due to fragmented datasets and isolated methods.

Standard ML:

Standard machine learning, although promising, tends to operate in isolation, relying on single-stage modeling and limited or fragmented data. This narrow scope makes it difficult to fully leverage available information, hindering comprehensive understanding and robust predictions.

Nature of Scientific Research:

Recognizing these challenges, our work generally aligns machine learning closely with the iterative, exploratory, and inherently multi-modal nature of real scientific research. We want to develope fully integrated, multi-stage machine learning pipeline, specifically designed to bridge the gap between theory, computation, and experiments. Our approaches aim to leverage diverse datasets, iterative refinement, and structured data comprehension—addressing previous limitations directly.

Future Directions:

Moreoever, we also would like to attempt to make improvement Represention the domain knowledge, most of existed work still use typical simplified manual encoding and treat in the ML modeling process. Our interest is also how to use LLM to process knowledge grpah, and potentially agent to automate the process to enable truly meaningful inverse design ability.