NATURE PODCAST
03 July 2019

Podcast: Machine learning in materials science, and sand’s sustainability

Listen to the latest from the world of science, with Benjamin Thompson and Nick Howe.

This week, using an algorithm to find properties in materials science, and the global consequences of sand-mining.

In this episode:

00:47 Predicting properties

A word-association algorithm is reading millions of abstracts to discover new properties of materials. Research article: Tshitoyan et al.; News and Views: Text mining facilitates materials discovery

08:28 Research Highlights

Tiny robot-jellyfish, and genome mutation hot-spots. Research Article: Multi-functional soft-bodied jellyfish-like swimming; Research Highlight: How DNA ‘hotspots’ snarl the search for cancer genes

10:48 Sand under strain

Researchers warn that the mining of sand is unsustainable. Comment: Time is running out for sand

15:44 News Chat

The results of a bullying survey, and the spread of microbial diseases through opioid use. News: Germany’s prestigious Max Planck Society conducts huge bullying survey; News: The US opioid epidemic is driving a spike in infectious diseases

Never miss an episode: Subscribe to the Nature Podcast on Apple Podcasts, Google Podcasts, Spotify or your favourite podcast app. Head here for the Nature Podcast RSS feed.

doi: https://doi.org/10.1038/d41586-019-02073-x

Transcript

This week, using an algorithm to find properties in materials science, and the global consequences of sand-mining.

Host: Nick Howe

Welcome back to the Nature Podcast. This week, we’ll be finding out how computers can predict the properties of materials…

Host: Benjamin Thompson

And hearing about the environmental impacts of sand mining. I’m Benjamin Thompson.

Host: Nick Howe

And I’m Nick Howe.

[Jingle]

Interviewer: Nick Howe

First up, I want to talk about making stuff. Materials science is a field all about creating the cutting edge, combining existing compounds and discovering new ones. But with so many different ways of going about it, it can be hard to know where to start. To discover an original material or even repurpose an old one with new properties requires a lot of experimentation with a little bit of luck. But this week in Nature, researchers have been applying a machine learning algorithm known as ‘Word2vec’ to try and streamline the discovery process of finding materials with new properties. This algorithm is known as unsupervised, meaning that it has no knowledge of what anything might mean – it discovers that through associations. In this case, the algorithm scanned millions of research paper abstracts to try and find meaning.

Interviewee: Vahe Tshitoyan

Well, we tried to see what the computer algorithm will learn about the different materials, applications and properties in the text. We just give it a simple task of looking at concurrences of different words in the text and that was the key idea – can a model that’s learned without any human input just by reading the text be useful for scientific discovery or scientific predictions in general?

Interviewer: Nick Howe

That was Vahe Tshitoyan, one of the authors of the new paper. The algorithm scanned the abstracts and looked at the eight words around a chosen term. The more often a word was found near this chosen term, the more likely it was to be associated.

Interviewee: Vahe Tshitoyan

The idea is that words that occur in similar contexts have similar meanings, so if a material is mentioned alongside similar application words and similar properties and similar synthesis techniques with another material, then these two materials are likely to be similar.

Interviewer: Nick Howe

In addition to looking for materials that might be similar, the algorithm could be used to look for new properties for existing materials. In Vahe’s case he was looking for materials that had thermoelectric properties – ones that can convert temperature to electricity and vice versa. To test the algorithm’s ability to predict materials that might have these properties, Vahe inputted abstracts from research papers published before 2009. When the algorithm found the word ‘thermoelectric’, it looked at what words were close by. Often these were chemicals or compounds. It then produced a list of top predictions for potential thermoelectric materials. Vahe then looked in the literature after 2009 to see if the algorithms top five predictions had come true.

Interviewee: Vahe Tshitoyan

It turns out that from these five predictions, three of them were studied as thermoelectrics in later years, and one of them is actually one of best present-day thermoelectric materials, so the algorithm was able to predict four, five, ten years in advance some of the good thermoelectrics that we use now.

Interviewer: Nick Howe

As this algorithm can quickly assess millions of abstracts and produce a short list of predictions about the properties of materials, Vahe hopes it will be able to be used as a scientific assistant, giving researchers some direction in what materials to study next. In fact, whilst he was writing this week’s Nature paper, some of the predictions the algorithm had made using the present-day literature came true.

Interviewee: Vahe Tshitoyan

We looked at the top 50 predictions, and three out of those 50 were studied while the paper was being reviewed and prepared for publication, so even in a short timescale, three of the predictions came true.

Interviewer: Nick Howe

Now, using machine learning in materials science is nothing new, but most of the time, the algorithms have worked on databases where different aspects of a material’s properties are encoded by a human. This means that these algorithms can only use what’s in the databases. There’s a whole world of information that they’re missing.

Interviewee: Olexandr Isayev

Much data in materials science are hidden in papers, in old catalogues, in handbooks, so it’s not available to work with.

Interviewer: Nick Howe

That’s Olexandr Isayev, a chemist who employs machine learning in his work, who’s also written a News and Views article on this topic. A database is only as useful as the data that’s put into it. By using written words rather than database information, Vahe’s algorithm can capture information hidden in things like papers and handbooks. Because of this, Olexandr thinks this word association algorithm is a useful method, but one that’s not without its limitations.

Interviewee: Olexandr Isayev

So, if you think about the words and the context of words, all of these have synonyms. They have slightly different meanings depending on the context. In many cases, for example, if the material has multiple properties, we are not really sure how good some of those methods would work.

Interviewer: Nick Howe

As this algorithm only discovers the meanings of words based on their frequency and proximity to other words, it only comes up with a single definition based on the occurrence of other terms. Basically, it struggles with synonyms. Another hurdle to overcome is that the algorithm would only be able to find connections for materials that have already been described. If there’s a new or hypothetical material not described in the literature, it wouldn’t be able to help. Vahe is confident that this a problem that can be solved by adjusting the algorithm to make predictions on new compounds based on the chemical structure of existing materials. He also thinks that future algorithms, that are better able to deal with complex sentences, could work on entire research papers, rather than just their abstracts. Vahe expects that reading the full text would allow the algorithm to make better predictions. Even so, he thinks the algorithm works well and is hopeful that in its current state, it can be used by other fields like drug discovery. As abstracts are free for everyone to read, Vahe thinks theoretically anyone could try this approach.

Interviewee: Vahe Tshitoyan

One thing that’s very exciting for me regarding this work is that all the data we used is freely available on the internet and also, the algorithms we used are currently very efficient and its very fast, like on a modern-day computer you can train a model in less than a day. So, in theory, anyone could do this research without having access to large computing power or any data that’s not available freely, which is something I hope this will spark. So, maybe younger scientists or people interested in machine learning and natural language processing, which might not necessarily have the resources, will be able to contribute to scientific progress.

Interviewer: Nick Howe

That was Vahe Tshitoyan of the Lawrence Berkeley National Laboratory in the US. You also heard from Olexandr Isayev of the University of North Carolina at Chapel Hill, also in the US. You can find Vahe’s paper along with Olexandr’s News and Views over at nature.com.

Host: Benjamin Thompson

The News Chat is coming up at the end of the show, where we’ll be hearing the results of a huge bullying survey. Coming up now, it’s time for the Research Highlights, read this week by Anna Nagle.

Anna Nagle

A team of researchers in Germany have created a tiny robot jellyfish just a few millimetres in size that – despite its diminutive diameter – is capable of performing a number of tasks. Although swimming robots aren’t new, scaling them down without losing functionality can be problematic. To overcome this, the researchers looked to juvenile jellyfish for inspiration, creating a robot with a magnetic core surrounded by eight tiny, flexible flaps. When an oscillating magnetic field is applied, the flaps contract and relax, propelling the bot along. By manipulating the flow of water around it, this miniscule machine can also lift small beads, bury itself and mix fluids. The researchers also say that their creation could be used to help understand environmental impacts on the biological jellyfish that were the inspiration for their robot’s design. You can find out more about that research in Nature Communications.

[Jingle]

Anna Nagle

Understanding the mutations in the human genome that can lead to cancer is an important area of work for many researchers. Tumours often contain multiple mutations, so to pinpoint the most important ones, scientists often focus on the DNA changes that occur most frequently. However, researchers in the US have found evidence of mutation-prone ‘hotspots’ that could foil attempts to identify the changes associated with tumour development. The team compared tumour genomes from more than 1,600 people to genome samples from their normal tissues, and found that certain sections of the genome fold into shapes that make them vulnerable to mutation, causing the hotspots. The new work suggests that some areas of the genome with multiple mutations may, in fact, not contribute to the disease, contrary to what was previously thought. Head over to Science to read more on that story.

[Jingle]

Host: Nick Howe

If I think of digging up sand, the first thing I think of is building a sandcastle. But the mining of sand is no child’s play. It’s a huge industry around the world, and its impacts are being felt by both people and the planet. In fact, humanity’s demand for sand is so large that it might outstrip supply by the middle of the century. A Comment piece in this week’s Nature sifts through this sandy problem, and one of its authors, Mette Bendixen, spoke with reporter Adam Levy.

Interviewer: Adam Levy

What is sand actually useful for?

Interviewee: Mette Bendixen

So, sand is the key ingredient actually in modern society. It’s used in buildings and construction and infrastructural projects. It is used in electronics, in glass, even in toothpaste and wine. So, we have a big use of sand globally.

Interviewer: Adam Levy

What is the actual scale of sand mining then?

Interviewee: Mette Bendixen

So, besides from water, sand is the natural resource we humans extract the most of – it even exceeds fossil fuels – so we’re talking around 30 to 50 billion tonnes that’s being used per year globally.

Interviewer: Adam Levy

It seems like, you know, we talk about fossil fuel extraction a lot, not just because of climate change but because of the impacts locally – why don’t we talk about sand extraction so much?

Interviewee: Mette Bendixen

Yeah, I’ve been wondering about that too. I think sand is something you take for granted. Sand is something you’re used to thinking is just everywhere. It’s in beaches, you have big deserts throughout the world. The thing about the deserts though is that you can’t use that as construction material because when the grains are being transported by wind, they’re simply too smooth and too rounded and they’re too well sorted. For the construction sand, that is not something that we can use unfortunately. The sand we use is coming from rivers, it’s coming from beaches.

Interviewer: Adam Levy

So, there’s this huge amount of sand being extracted. What are the actual impacts of this mining?

Interviewee: Mette Bendixen

It affects both nature and people. In South Asia, we see that ecosystems are being destroyed. River banks are collapsing which means that people lose their houses as they simply fall into the river, and that means that people have to move, and the Vietnamese government has estimated that up to a half a million people will need to move away from the flood plains and the river banks of the Mekong delta within the near future.

Interviewer: Adam Levy

But the way sand is being extracted is also causing big problems.

Interviewee: Mette Bendixen

Yes, a lot of the sand that is mined today is mined illegally. It’s being controlled by actual ‘sand mafias’. In Kenya, kids drop out of school to mine sand.

Interviewer: Adam Levy

You’re also looking at ways in which this problem could potentially be fixed. How would we begin to get a handle on sand mining?

Interviewee: Mette Bendixen

Yeah, so we say that the solutions lie in the alternatives, and that means alternative technologies. Just last year, the Imperial College London, they were actually able to make this concrete-like material with sand from the desert and with half the CO₂ footprint. I think what is important also to do is reuse, which can be done when you destroy older buildings and simply reuse the material here so It’s not just being lost.

Interviewer: Adam Levy

So, how do we get from these recommendations to actually implementing something for the world that would reduce this pressure on sand?

Interviewee: Mette Bendixen

We call upon the United Nations to establish a global monitoring programme for sand resources because really, we need to understand how much sand do we have, where is this sand and how much of it is being extracted. Right now, we don’t have that. We don’t have that global overview, so we cannot use this material in a sustainable way. And I feel like the problem of sand scarcity is getting more attention right now, so I feel like we’re towards creating a momentum where this will get more focus.

Interviewer: Adam LevyHow do you actually feel about whether we can and we will turn this problem around?

Interviewee: Mette Bendixen

I think there is becoming an increasing awareness that we need to do something about this issue because sand is not something we should take for granted, and after having worked with this, I realise now how much sand is actually being used. Just outside my door, they were redoing the asphalt there and I realised well, that’s sand there, just across my building from where I sit there’s being built a new building. There’s just sand everywhere and you don’t think of that.

Host: Nick Howe

That was Mette Bendixen talking to reporter Adam Levy. Catch her Comment piece over at nature.com/opinion.