This post is a (personal) summary of the important DOE report AI for Science that appeared in February 2020. Though the official title is “for Science”, there are important details concerning “engineering” too.

The DOE defines AI for Science as follows:

“AI for Science” broadly represents the next generation of methods and scientific opportunities in computing, including the development and application of AI methods (e.g., machine learning, deep learning, statistical methods, data analytics, automated control, and related areas) to build models from data and to use these models alone or in conjunction with simulation and scalable computing to advance scientific research.

I would add: “… to advance scientific research and engineering.” But this approach is precisely what we have defined as Digital Twins, in particular with the “in conjuction” descriptor.

What are the major conclusions?

  1. Learned models will replace data.
  2. Experimental discovery will be refactored.
  3. Questions will be pursued semi-autonomously.
  4. Simulation and AI will merge.
  5. Theory will become data.
  6. AI laboratories will be created.

These will all be gradually transferred to the social, medical, environmental and engineering worlds. In the book some of these points are addressed in Chapters 13 and 14. In particular there is a discussion of the “robot scientist”, related to point 7 above, and its implication in terms of abductive reasoning (points 3 and 5).

Additional references

  1. At the recent NeurIPS 2021 conference, a workshop dedicated to AI for Science, entitled, Mind the Gaps.
  2. The question of ethical, responsible use of AI is an emerging and vital topic. Stuart Russell just gave a series of fascinating lectures on the subject, entitled “Living with Artificial Intelligence”.
  3. A growing number of papers appearing in Nature and Science. Nature now has its own AI journal, Machine Intelligence, with monthly editions that include an excellent selection of papers.