An AI To Identify The Environment A Grain Of Sand Came From
8:48 minutes
If you were given a bucket of sand and asked to determine where it came from, you’d probably have a hard time guessing if it was from a beach, a riverbank, the playground down the street, or a Saharan sand dune.
There are experts who can make a guess at that sort of ID, using a categorization process that takes skill, a scanning electron microscope, and hours of time. Now, however, researchers report in the Proceedings of the National Academy of Sciences that they’ve developed an AI model that can quickly judge whether a sample of sand came from a beach, a river, a glacial deposit, or a wind-blown dune.
That type of identification isn’t just of interest to geologists. Sand is one of the world’s most in-demand resources, second only to water in use. And different applications need different types of sand—for instance, making concrete and mortar requires angular sand for good adhesion and stability. These kinds of needs have given rise to illicit sand mining, sand theft, and sand smuggling. A way of rapidly identifying the origins of a sample of sand could be useful to investigators, or to companies seeking to ensure sustainability goals.
Michael Hasson, a PhD candidate in Stanford’s Department of Earth and Planetary Sciences, joins SciFri’s Charles Bergquist to discuss the new SandAI, and the challenges of tracking grains of sand.
Keep up with the week’s essential science news headlines, plus stories that offer extra joy and awe.
Michael Hasson is a Ph.D. candidate in the Department of Earth and Planetary Sciences at Stanford University in Stanford, California.
SPEAKER: For the rest of the hour, Sci Fri’s Charles Bergquist is here with a story about sand. Hi, Charles.
CHARLES BERGQUIST: Hey, Ira. If I gave you a bucket of sand and asked you to tell me where it came from, I’m guessing you’d have a hard time telling me if it was from a Saharan sand dune, an ocean beach or a riverbank. There are experts who could make a guess at that sort of ID, based on a process that takes skill, a scanning electron microscope and hours of time. But of course, now, there’s an AI for that.
Joining me now is Michael Hasson. He’s a PhD candidate in Stanford’s Department of Earth and Planetary Sciences and lead author on a paper describing this new AI for sand that was published this week in the proceedings of the National Academy of Sciences. Welcome to Science Friday.
MICHAEL HASSON: Hey, thank you so much for having me. It’s great to be here.
CHARLES BERGQUIST: Great to have you. Walk me through this challenge I mentioned. I give you a bucket of sand, how would you start trying to ID it without your tool?
MICHAEL HASSON: Yeah, so in the past, it’s basically amounted to filling out a spreadsheet with about 30-35 different features. And a human would go through each photo and make a check mark next to each feature that was present. Things like whether the grain is angular or subangular or rounded, the presence of arcuate steps, parallel striations or elongated depressions. And so at each step, the human is making this judgment call about very specific terms that have really big implications for where the sand actually comes from. So it’s a hugely subjective process and it takes a really long time.
CHARLES BERGQUIST: Is this something that a regular person could pull out a magnifying glass and see some of this, or do you really need your scanning electron microscope to make this determination?
MICHAEL HASSON: Yeah, we really need a microscope to do this, and particularly, the scanning electron microscope that is able to resolve these features that are less than a micron.
CHARLES BERGQUIST: Is there any kind of standard for categorizing sand? Do sand people look at a sample and say, “oh, that’s clearly a Josephson type 2A”?
MICHAEL HASSON: No, it really all exists along a spectrum. And so using some statistical analysis, we can tell you if it’s more like one environment or another, but we can’t really give you hard, specific answers about where it exactly comes from.
CHARLES BERGQUIST: So forgive me for asking this, but is there really a huge need for rapid, accurate sand identification?
MICHAEL HASSON: Yeah, that’s a totally fair question. Outside of fundamental sciences like geology, geoarchaeology, something that the world is going to start having to grapple with more and more is where we get our sand from. Sand sourcing is this really critical problem because after water, it’s the second most important raw material in the world. And anything that we build that’s made out of concrete is reliant on sand, and not just any sand. It has to be angular sand so that the grains lock together and form a structurally sound material.
And so what that’s led to is a lot of illegal mining of sand, especially in places like southeast Asia that are developing very rapidly. They’re often companies that set up on a riverbank and they just start removing sediment. And so if you’re a construction company that’s trying to make sure you’re sourcing your sand responsibly, using our tool is a way that you can confirm whether your supplier is being accurate about what they’re telling you and where that sand is coming from.
CHARLES BERGQUIST: Yeah. So break it down for me. How structurally is sand from, say, a beach dune different from sand on a Saharan desert dune to an expert?
MICHAEL HASSON: Different environments can produce very similar textures. So it’s really helpful to look at these in ensemble. And so when a human has been doing this in the past, they’ve taken the sample of, say, 100 grains and they say, OK, on average, this sample is a little bit more round than another. And if it’s a little bit more round, maybe that means it was moved by wind instead of water.
Because when grains collide in the air, there’s less of a viscous fluid to soften the impact between them. And so when they collide in the air, they just break off and the grains become more round. But it’s a non-unique process. And so that’s why we were really excited to apply machine learning to this, to try to capture these patterns that are really difficult to get at from just a human analysis.
CHARLES BERGQUIST: So now you have this AI tool that you’ve trained on thousands of known samples of sand. Can it do this task better?
MICHAEL HASSON: Yeah. So this tool that we’ve developed doesn’t just take one sample and condense it down to a single result. So historically, we’ve taken this sample of, say, 100 sand grains, and that sample is more or less beach. That would be the answer that you could get from it. With this new tool, we can get an answer for every single sand grain, which lets us do a much better job of fingerprinting individual sites and telling more complex stories of whole samples. And so one really neat example of this was from a sample that we used to test the model from the bottom of the Mississippi river.
So the Mississippi river is about 2,000 miles long, and we sampled from very end of it. And rather than showing an entirely river dominated signature, the sample was about 50% river and 50% windblown sand. And so at first, that might seem a little bit puzzling, but it’s telling this really interesting story.
And the story is that way upstream, the banks of the Mississippi river are made out of windblown sediment that was deposited around the last ice age. And so just using our tool, we can start to piece together the story of a modern river moving across this bed of ancient windblown sand and get this really interesting picture of how that part of the world has evolved just by looking at a few grains of sand.
CHARLES BERGQUIST: So if you come up with this answer that’s, like, this is 50% windblown, 50% river, does that give you a fingerprint of the Mississippi? If you come up with those numbers, do you say, this is probably Mississippi river sand, as opposed to some other place?
MICHAEL HASSON: Right now, we’re working with probably. We haven’t compiled that whole data set of what different major rivers, let’s say, around the world would produce in terms of the distribution of different environments that our model would predict. But that’s likely a really good signature for individual sites that we could use going forward.
CHARLES BERGQUIST: I mean, if you were able to give it enough samples of very specific rivers, could you fine tune this model to be even more specific, like picking out one river environment versus another?
MICHAEL HASSON: Yeah, absolutely. So that brings up this really important point within the broader machine learning world, which is the idea of transfer learning. And transfer learning is this idea that a model that’s trained to do one task well can likely be fine tuned to do another similar task even better than it would be able to do if it were just starting from scratch. And so what we’ve achieved now is we have a model that sort of knows how to work with sand.
And if we were to take this model and compile a different set of data where we know exactly what site the sample comes from rather than what environment it is made up of, we could train the model to distinguish between different rivers, let’s say, and it would be expected to perform much better than if we were just training the model from scratch.
CHARLES BERGQUIST: I mean, you mentioned sand forensics in the context of construction companies. But is there any possible use here for regular police forensics or, I don’t know, archeological exploration or something like that?
MICHAEL HASSON: It’s something that people have looked into for regular police forensics in terms of using it to test a sample that was on a murder suspect shoe and seeing if it confirms or refutes their alibi. Geoarchaeology is a context that we’re really excited for this to be used in too because this is a new tool for contextualizing environments when we have really limited data available.
One really neat example that we found had to do with taking a core through an archeological site with the goal of contextualizing where artifacts were coming from. So where were people living? Was this a river? Where they were living in a desert? And that’s really hard to do, especially when there’s no real rock outcrop that geologists could look at. And so when you take a core to minimize the destruction of the site, you also lose all the context to where that sediment is coming from. And so having a tool like this that can provide some environmental information without any context can be really critical in some situations.
CHARLES BERGQUIST: So where do you go from here with this? What would be the next step in this research?
MICHAEL HASSON: So we’ve publicized the model so that it’s online for anybody to use freely. And we’re hoping that as more and more people use it, we will understand better the instances where it works and the instances where it doesn’t. We’ve tested it to know that it works quite well in as much data as we have available. But given the natural variability of the world, there will almost certainly be places where it fails. And knowing about those will help us tweak the model and come up with a new version of it that will work in an even wider variety of settings.
CHARLES BERGQUIST: Very cool. Michael Hasson is a PhD candidate in Stanford’s department of Earth and Planetary Sciences. Thanks so much for taking time to talk with me today.
Copyright © 2024 Science Friday Initiative. All rights reserved. Science Friday transcripts are produced on a tight deadline by 3Play Media. Fidelity to the original aired/published audio or video file might vary, and text might be updated or amended in the future. For the authoritative record of Science Friday’s programming, please visit the original aired/published recording. For terms of use and more information, visit our policies pages at http://www.sciencefriday.com/about/policies/
As Science Friday’s director and senior producer, Charles Bergquist channels the chaos of a live production studio into something sounding like a radio program. Favorite topics include planetary sciences, chemistry, materials, and shiny things with blinking lights.
Ira Flatow is the host and executive producer of Science Friday. His green thumb has revived many an office plant at death’s door.