I just read an Eureka Alert (see also ZDNet’s blog)mentioning that a couple of researchers at MIT found new, potentially useful crystal structures with AI and Data Mining techniques. You can find the abstract of their paper here. I’ve seen randomness and Genetic Algorithms around alot lately (such as the self-reconfigurable-modular-robot/) and a robot that can do bioinformatics experiments (DNA sequencing) all by himself (link?). I think that this is a very useful application of AI. However, it is only an application of the scientific knowledge. It’s fast testing based on the current physical models and insights. It automates science to an extend, but does not come up with new insights. It’s more data without more people to add an interpretation. For example, it took a few years before somebody found an application for Teflon.
I haven’t seen this around (will search again), but what would be really interesting is an algorithm that can form a new hypothesis (e.g. a differential equation) based on outcomes from Physics experiments. An algorithms that explains the data and forms a theory. It’s probably harder to build than regression algorithms…