Jason from 404 Media wrote about a paper written by Anthony K. Cheetham and Ram Seshadri published in Chemistry of Materials which found that in the 380000 materials out of the 2.2 million ones discovered by Google’s DeepMind they didn’t find “any strikingly novel compounds”.
This is disappointing because AI generated SEO spam has already ruined the search experience on the internet. The focus of those websites is quantity over quality and to rake in as much advertising revenue as possible. Academic papers are the last place you want this enshittification to creep in as they reflect humanity’s scientific progress. (Or maybe we are too late for that?). There is no point in having a huge database of materials that contain mostly imaginary compounds or stuff that can’t practically exist.
One of the most bizarre things in the article is that DeepMind predicted H2O11 would exist and is somehow useful. I assumed that companies with resources like Google and DeepMind would probably have some checks and balances in place to modify weights and biases on their models to work within basic constraints of chemistry. ML is used extensively in Engineering. Algorithms are used in simple applications such as curve fitting methods to difficult tasks like prediction of wear in complex multi-component systems. There is actual research being done in this field for the betterment of humanity and flashy research papers like these and the superconductor fiasco ones don’t help but instead erode trust.