$data->pivot->title

Machine-learning serving science

Scientific research is inspired by the real-world problems and up-to-date challenges addressed in timely communication and collaboration. All around the world research results are published in scientific journals, as the information exchange is a crucial aspect of progress and development. As such, scientific journals are valuable research and study material, containing correlations that can lead to future discoveries if the information is combined in a relevant manner. Unfortunately, due to this vast number of articles, it is truly impossible for a scientist to browse through all relevant literature just to extract necessary data required to conduct an out-and-out study. The majority of scientific discoveries remain under the radar. This is a tremendous problem for researchers who are not in the position to promote their results outside of a narrow circle of colleagues.  

"Scientific progress relies on the efficient assimilation of existing knowledge in order to choose the most promising way forward and to minimize re-invention."

V. Tshitoyan et al., July 2019 in Nature

A group of researchers from Berkeley university might have found a solution to this problem. They constructed a machine-learning model to extract textual correlations from scientific publications, and conduct a focused search for a group of previously unnoticed thermoelectric materials. They used about 3.3 million scientific abstracts that addressed physico-chemical properties of various classes of materials, not necessarily foreseen in the investigated domain. As a result, a new class of thermoelectric materials was discovered. This impressive example of tangling research results from various science fields is published in Nature magazine in July. The importance of this discovery and impact of ML model can be easily understood, having in mind that the overall commercialization of thermoelectric effect is still limited due to low efficiency. This new approach to mining and reviewing scientific data can significantly speed up the process of finding adequate materials, globalizing usage of the thermoelectric effect, and ultimately formulating novel ideas for further studies.