Nowadays, machine learning seems to exist at every corner of scientific studies. When I walk into a conference, skim through journals, or tune into a science podcast, I can find the stirred excitement about this trendy topic. A huge amount of funding from the Department of Science was just poured in to advance the discovery in chemistry and material sciences. I couldn’t help but wonder, since when this fascination creeps its way across fields of research. Most relevantly, do I even have to care about it while I’m not studying algorithm or programming at all?
Besides googling a bunch of definition and application of machine learning, I searched on Web of Science for publications. As of August 2019, the web of science database includes over 25,000 published articles having ‘machine learning’ and ‘materials.’ The number of publications burst around 2017 and is expected to reach a new high by the end of this year.
In one word, materials research is on the front wave, among others, in the rise of machine learning development.
Figure. The number of publications was generated from a Web of Science search for publications that include ‘machine learning’ and ‘materials.’ (data retrieved at 08/08/2019)
I looked at the titles on the top-cited publications. Along with my encountering experiences, the element of machine learning in these materials studies plays out with a certain pattern. After collecting the chemical and physical properties, human researchers feed the information to build a digital dataset. Using the dataset, the machine learning method can classify those input and predict the properties, purely based on statistics. Because the human brain cannot, simply put, comprehend the overwhelming number of combinations of information, we need machines, or computers, to process the data.
This is all great. A faster and more efficient tool is essential for technology breakthrough. But again, how do I appreciate the hot topic if my research barely has to do with it?
I refined the journals published since 2018 to understand the trend, which led me to a well-written and easy-to-follow reviewing article, “Machine learning for molecular and materials science,” published by Nature. The paper envisioned a promising future, as always, machine learning holds. The most impressing part with it is that the key to revolutionizing molecular and material discovery is the accessible tools for non-specialists.
In materials science, the chemical-simulation toolkits offer researchers to anticipate the properties of a compound. Over the years, there are open-source tools developed to perform the structure-property relationship for molecules and materials. Just as those toolkit software made simulation accessible to users with little or no theoretical training, machine learning techniques can broaden the routine examination by transforming the traditional research workflow.
This perspective strikes me. As an experimentalist, I am used to feeling that I’m out of the bubble of computation or programming development, while I have already been involved in the evolution of research workflow. In the past, I have learned to implement computational chemistry software to support the experimental observations. This approach, by the authors’ definition, is a first-generation workflow which relies on prior knowledge of structure. In the future, I could be using one of machine learning-based toolkits to boost the understanding and extend the influences.
As soon as I realize the waves I’m standing on, I’m more confident in embracing the new and unknown, rather than overlook or even fear it. It might take me a while to fully comprehend the principles yet I’m assured that I don’t necessarily have to master in algorithms. The most important mindset is to keep looking out for learning and appreciating beyond the newly rising power.
Comments