Automated Discovery

[Image: BuySellGraphic.com]

In 1986, Dan Swanson from the University of Chicago discovered a discovery.

Swanson (who passed away in 2012) was an information scientist and a pioneer in literature analysis. In the 1980s, he studied the distribution of references in scientific papers and found that, on occasion, studies on two separate research topics would have few references between them, but would refer to a common, third, set of papers. He conjectured that might indicate so-far unknown links between the separate research topics.

Indeed, Swanson found a concrete example for such a link. Already in the 1980s, scientists knew that certain types of fish oils benefit blood composition and blood vessels. So there was one body of literature linking circulatory health to fish oil. They had also found, in another line of research, that patients with Raynaud’s disease do better if their circulatory health improves. This led Swanson to conjecture that patients with Raynaud’s disease could benefit from fish oil. In 1993, a clinical trial demonstrated that this hypothesis was correct.

You may find this rather obvious. I would agree it’s not a groundbreaking insight, but this isn’t the point. The point is that the scientific community missed this obvious insight. It was right there, in front of their eyes, but no one noticed.

30 years after Swanson’s seminal paper, we have more data than ever about scientific publications. And just the other week, Nature published a new example for what you can do with it.

In the new paper, a group of researchers from California studied the materials science literature. They did not, like Swanson, look for relations between research studies by using citations, but they did a (much more computationally intensive) word-analysis of paper abstracts (not unlike the one we did in our paper). This analysis serves to identify the most relevant words associated with a manuscript, and to find relations between these words.

Previous studies have shown that words, treated as vectors in a high-dimensional space, can be added and subtracted. The most famous example is that the combination “King – Man + Woman” gives a new vector that turns out to be associated with the word “Queen”. In the new paper, the authors report finding similar examples in the materials science literature, such as “ferromagnetic −  NiFe + IrMn” which adds together to “antiferromagnetic”.

Even more remarkable though, they noticed that a number of materials whose names are close to the word “thermoelectric” were never actually mentioned together with the word “thermoelectric” in any paper’s abstract. This suggests, so the authors claim, that these materials may be thermoelectric, but so-far no one has noticed.

They have tested how well this works by making back-dated predictions for the discovery of new thermoelectric materials using only papers published until one of the years between 2001 and 2018. For each of these historical datasets, they used the relations between words in the abstracts to predict 50 thermoelectrical materials most likely to be found in the future. And it worked! In the five years after the historical data-cut, the identified materials were on average eight times more likely to be studied as thermoelectrics than were randomly chosen unstudied materials. The authors have now also made real predictions for new thermoelectric materials. We will see in the coming years how those pan out.

I think that analyses like this have lot of potential. Indeed, one of the things that keeps me up at night is the possibility that we might already have all the knowledge necessary to make progress in the foundations of physics, we just haven’t connected the dots. Smart tools to help scientists decide what papers to pay attention to could greatly aid knowledge discovery.