Text mining facilitates materials discovery | Olexandr Isayev, Nature, Vol. 571, 4 July 2019
Materials informatics uses machine learning to discover new materials with desirable properties. This field used both supervised and unsupervised approaches.
This short paper concerns the role text mining has to play in this field.
Tshitoyan et al. created a curated dataset of 1.5 million abstracts on materials science, physics and chemistry. They then piped this through word2vec (unsupervised) to create a high-dimensional representation of the word embeddings in this field.
The resulting word vectors are called embeddings, and are used to position each word, represented as a data point, in a multidimensional space that represents the vocabulary. Words that share common meanings form clusters within that space. (42)
Succes story: the clusters found using word embeddings reflected known chemical concepts.
The researchers found that the obtained word embeddings for materials-science terms produced word associations that reflect rules of chemistry, even though the algorithm did not use any specific labels to identify or interpret chemical concepts. When combined using various mathematical operations, the embeddings identified word associations that corresponded to concepts such as ‘chemical elements’, ‘oxides’, ‘crystal structures’, and so on. The embeddings also identified clusters of known materials (Fig. 1) corresponding to categorizations that could be used to classify new materials made in the future. (43)
But this also means we may leverage this correspondence for materials discovery. In this case, it was used to 'discover' new properties of existing materials.
If a report on a material is highly similar (in the high dimensional space) to reports concerning thermoelectric materials, can we then predict with above random likelihood that this material also has thermoelectric properties?
The authors validated this approach by training a model using literature published before a particular cutoff year, and then checking to see whether it identified materials that were reported to be thermoelectric in subsequent years. The top 50 materials picked using this method were 8 times as likely to have been studied as a thermoelectric in the 5 years after they had been reported than were randomly chosen materials. (43)
A clear downside of this approach is that it cannot be used to help find materials that are not yet described in the literature.