Data Backup Digest

Highly advanced artificial intelligence (AI) is increasingly making its way into mainstream society. From popular video games to next-gen software and even AI-powered robotics, it certainly seems that we're heading into a future that relies heavily on AI.

But AI is useful in scenarios other than automation and decision-making – it seems the latest AI systems are capable of performing scientific research without human guidance or supervision. Moreover, these systems are intelligent enough to make brand new scientific discoveries that have been missed by human researchers and scientists throughout all these years.

The AI system in question is the product of a team of developers and researchers with the Lawrence Berkeley National Laboratory in Berkeley, California. It analyzed the language used in millions of archived scientific papers – using an advanced algorithm known as Word2Vec – to ultimately make new connections and elicit scientific discovers that were previously unknown.

One such discovery involved the potential for new thermoelectric materials, which could be used in various heating and cooling installations. Moreover, the AI system was able to make this prediction without even knowing anything about thermoelectricity or materials science. Instead, it came to this conclusion using nothing more than word associations. Not only did it highlight several potential replacements for our current thermoelectric materials, but some might even perform better than the ones we currently use.

Anubhav Jain, a researcher with the Lawrence Berkeley National Laboratory, summed it up by saying: "It can read any paper on material science, so can make connections that no scientists could. Sometimes it does what a researcher would do; other times it makes these cross-discipline associations."

Jain also went into detail about the functionality of Word2Vec, saying: "The way that this Word2vec algorithm works is that you train a neural network model to remove each word and predict what the words next to it will be. By training a neural network on a word, you get representations of words that can actually confer knowledge."

The neural network in question contained approximately 3.3 million scientific abstracts, most of which directly pertained to materials science. The algorithm used this to develop a vocabulary containing approximately 500,000 different words, which were then entered into Word2Vec for relationship analysis.

It was this 500,000-word vocabulary that resulted in the breakthrough. Additionally, the AI system was able to gain a comprehensive understanding of the periodic table of elements and the chemical structure of molecules – all from the examination of old scientific documents.

However, the final results have yet to be proven. As such, the research team with the Lawrence Berkeley National Laboratory wanted to see if the algorithm could actually predict a scientific breakthrough before it happened. To achieve this, they analyze papers before 2009 – and were amazed to see that their AI system would have successfully predicted one of today's best thermoelectric materials all the way back in 2008. As it is, the material in question wasn't discovered until 2012.

The implications of such a highly advanced AI system are great. Not only does it have the potential to spot errors or missed connections in past scientific research, but it can quicken the pace of new scientific discoveries in the future.