Can AI help science? In my opinion, yes, and it can do so in several directions. This week, I found three interesting examples that showcase the potential of AI to foster creativity and productivity in science.
1 - AI can help write equations that push scientific discovery further.
AI Hilbert is an algorithm that can generate and derive new equations from a combination of existing theory and data, intending to fill in gaps in scientific knowledge, including introducing new theories. The tool is based on multivariate polynomial expressions and is aimed at physics theory. The main goal is to augment the scientific method in contexts where collecting data can be expensive or noisy, data-driven models can be overly rigid, and where "traditional" AI would generate equations that can be poorly generalizable.
This is a significant advancement for areas where scientific discovery has been stagnating (See, for example, research from @Prof.Arora and co-authors: https://onlinelibrary.wiley.com/doi/epdf/10.1002/smj.2693).
IBM technical blog (https://bit.ly/4gajYn3)
Nature Communication paper: https://www.nature.com/articles/s41467-024-50074-w
2- AI can help with scientific literature understanding.
Scientists build their knowledge on existing work, and reading the existing scientific literature is a fundamental step. There are several ready-to-use AI-based tools on the market, but the common problem in using AI for literature reviews arises when semantics for each discipline can be very different, specific jargon may not be captured by LLMs, and researchers receive hallucinated answers.
SciLitLLM is a suite of LLMs that presents promising results in understanding specialized scientific literature. It is built on a hybrid approach that combines continual pre-training (CPT) and supervised fine-tuning (SFT) to help adapt to different domains and infuse instructions for domain-specific tasks.
This tool can help scientists extract targeted information, summarize a large corpus from a stack of documents, and speed up their literature searches. However, it won't substitute for the researchers' critical reading and understanding of such knowledge.
Pre-print: https://arxiv.org/abs/2408.15545v1
3- AI can help generate novel research ideas
Researchers from Stanford University did an experiment with 100+NLP researchers.
Pre-print: https://www.arxiv.org/abs/2409.04109
While I am very excited to read about these novel ways AI can help scientists produce innovation, there are some caveats. These examples are not plug-and-play tools that researchers can use directly for their jobs and require some programming skills to be customized to the scientists' needs. It looks like it still requires time, trust from the scientific community, and investments from the institutions to make these inventions ready to be consumed by the majority of scientists.
For those working in research and science, I am curious to hear your opinion!