The Artificial Geologist: Using Machine Learning & Neural Networks to Find Gold

WEBVTT
00:00:11.020 --> 00:00:16.899
Lost in the fog of a cosmic storm Floating on
00:00:16.899 --> 00:00:21.140
whimsical wavelengths is the norm Dancing through
00:00:21.140 --> 00:00:30.529
the stars, chasing spectrums of light Welcome
00:00:30.529 --> 00:00:33.530
back to Whimsical Wavelengths. This time we're
00:00:33.530 --> 00:00:36.170
digging deep into the bedrock of discovery. This
00:00:36.170 --> 00:00:38.789
episode is where geology meets algorithms, and
00:00:38.789 --> 00:00:41.250
the motherlode is just a model away. We're using
00:00:41.250 --> 00:00:43.310
computers to crunch the crust, letting machine
00:00:43.310 --> 00:00:45.350
learning mine the patterns hidden in the data.
00:00:46.170 --> 00:00:49.310
This episode is loaded with knowledge that's
00:00:49.310 --> 00:00:52.149
anything but artificial, even though the intelligence
00:00:52.149 --> 00:00:57.009
is... And the puns are terrible. So machine learning...
00:00:57.179 --> 00:00:59.179
The topic of this episode is a bit outside of
00:00:59.179 --> 00:01:01.460
my wheelhouse, so I have a guest, and no, not
00:01:01.460 --> 00:01:04.459
just any guest, a co -worker from my alter ego.
00:01:04.680 --> 00:01:07.260
Yes, so far we've had a couple of guests that
00:01:07.260 --> 00:01:09.560
use scientific results as the basis for their
00:01:09.560 --> 00:01:12.799
career outside of academic institutions. Let's
00:01:12.799 --> 00:01:16.040
add one more. Before I get further into the introduction,
00:01:16.260 --> 00:01:17.980
many months ago I had asked for someone from
00:01:17.980 --> 00:01:20.579
my company I work for, Computational Geoscience,
00:01:20.680 --> 00:01:23.420
to come on the podcast and talk about machine
00:01:23.420 --> 00:01:25.980
learning and its applications in mineral exploration.
00:01:26.780 --> 00:01:29.299
It got delayed due to getting some results and
00:01:29.299 --> 00:01:32.599
a shuffle of personnel and more normal real life
00:01:32.599 --> 00:01:35.560
stuff. So this conversation is months in the
00:01:35.560 --> 00:01:39.200
making. It's also all about science. Lots of
00:01:39.200 --> 00:01:41.560
science takes place within the corporate world.
00:01:41.640 --> 00:01:44.019
This here is a chance to look at one aspect.
00:01:44.299 --> 00:01:46.920
It isn't all about laboratories and classrooms.
00:01:47.099 --> 00:01:49.879
It's also about boardrooms. There's lots of ways
00:01:49.879 --> 00:01:54.719
to have a career in science. Now, let's set it
00:01:54.719 --> 00:01:57.010
up. We've been mining minerals from the ground
00:01:57.010 --> 00:01:59.569
for thousands of years. All the deposits close
00:01:59.569 --> 00:02:01.730
to the surface and near people have been mined.
00:02:01.730 --> 00:02:05.969
They're gone. So the deposits we have left are
00:02:05.969 --> 00:02:10.650
either in remote places or many hundreds of meters
00:02:10.650 --> 00:02:13.509
under the ground. That means they're harder to
00:02:13.509 --> 00:02:16.770
find and harder to extract those shiny resources
00:02:16.770 --> 00:02:20.689
needed for our modern world. Today we also have
00:02:20.689 --> 00:02:23.870
a lot of data. satellite imagery topography lidar
00:02:23.870 --> 00:02:26.370
geophysical surveys like gravity see episode
00:02:26.370 --> 00:02:29.490
12 for an explanation on boogie gravity or magnetics
00:02:29.490 --> 00:02:32.669
there's so much more you'll see that sites that
00:02:32.669 --> 00:02:35.430
have a lot of data typically also have a lot
00:02:35.430 --> 00:02:37.870
of drilling this is where they've cored into
00:02:37.870 --> 00:02:41.389
the earth to take samples deep within i think
00:02:41.389 --> 00:02:43.569
this is a good detour drilling especially when
00:02:43.569 --> 00:02:46.210
the area is not near a road it's it's quite an
00:02:46.210 --> 00:02:47.979
endeavor The drill rig is something that needs
00:02:47.979 --> 00:02:51.840
to be moved by heavy machinery or a large truck.
00:02:52.139 --> 00:02:54.840
Flying it into sight is not simple in many areas.
00:02:55.280 --> 00:02:57.719
They take a lot of fuel, a crew of at least four
00:02:57.719 --> 00:03:01.439
people to run 24 -7. Each hole, after all is
00:03:01.439 --> 00:03:04.780
said and done, costs well over $100 ,000. And
00:03:04.780 --> 00:03:07.520
what do you get? Cylindrical sections of rock
00:03:07.520 --> 00:03:11.719
for the entire drill hole. Being approximately
00:03:11.719 --> 00:03:13.900
six centimeters in diameter, the information
00:03:13.900 --> 00:03:18.159
brought up is expensive. It's the only way to
00:03:18.159 --> 00:03:20.439
directly observe what we infer from the surface.
00:03:20.620 --> 00:03:23.699
Is there mineralization there? Will it be there
00:03:23.699 --> 00:03:28.159
next? Mine. A lot is done with that core of rock.
00:03:28.520 --> 00:03:31.319
Geochemistry, rock types, structures, like fractures,
00:03:31.319 --> 00:03:34.060
are all written down and described. The point
00:03:34.060 --> 00:03:37.699
is you don't want to miss with the drill. Drilling
00:03:37.699 --> 00:03:41.319
is expensive. Whatever increases your success
00:03:41.319 --> 00:03:45.039
rate and costs less than drilling is useful.
00:03:46.370 --> 00:03:50.490
Enter mineral prospectivity. Geophysics, drilling,
00:03:50.610 --> 00:03:54.229
remote sensing. With all that data, it is still
00:03:54.229 --> 00:03:58.490
hard to find the shiny stuff. And as said, drilling
00:03:58.490 --> 00:04:01.889
is expensive. To help combat this, we turn to
00:04:01.889 --> 00:04:03.830
machine learning and artificial intelligence
00:04:03.830 --> 00:04:06.189
to find patterns in the data humans may miss.
00:04:06.370 --> 00:04:10.069
To get enlightenment on this, a data scientist
00:04:10.069 --> 00:04:13.210
and my co -worker from the best place to work
00:04:13.210 --> 00:04:16.800
on Earth. And no. I'm not just saying that. He
00:04:16.800 --> 00:04:18.899
joins the show. Here with us today is a data
00:04:18.899 --> 00:04:21.860
scientist with Computational Geosciences Incorporated.
00:04:21.959 --> 00:04:24.300
He got his Bachelor's of Science from the University
00:04:24.300 --> 00:04:26.500
of Leeds, a Master's from Western University,
00:04:26.720 --> 00:04:29.160
and then transitioned to industry using his degrees
00:04:29.160 --> 00:04:31.660
in geophysics and continues to use science to
00:04:31.660 --> 00:04:34.759
solve problems. Please welcome Frederick Jackson
00:04:34.759 --> 00:04:38.000
to Whimsical Wavelengths. Hi, Jeff. Thank you
00:04:38.000 --> 00:04:40.620
for having me. I'm a big fan. Oh, that's nice.
00:04:41.069 --> 00:04:44.629
And no, I didn't pay him to say that. For those
00:04:44.629 --> 00:04:46.910
that are longtime listeners, the reason why I
00:04:46.910 --> 00:04:48.610
made that joke is because the listener question
00:04:48.610 --> 00:04:51.250
at the end of episode 11 about the Spinosaurus
00:04:51.250 --> 00:04:54.629
controversies actually came from Frederick here.
00:04:54.850 --> 00:04:58.129
A trick up my sleeve to make sure at least one
00:04:58.129 --> 00:05:02.449
more person listened to the episode. But more
00:05:02.449 --> 00:05:04.509
broadly, how does your interest in dinosaurs
00:05:04.509 --> 00:05:07.189
and other science avenues fit into your path
00:05:07.189 --> 00:05:10.110
to where you are today as a data scientist? Yeah.
00:05:10.170 --> 00:05:14.110
And by the way, I follow paleontology quite closely.
00:05:14.170 --> 00:05:16.279
And there was actually... I don't know when you'll
00:05:16.279 --> 00:05:18.220
release this episode, but at the time of recording,
00:05:18.360 --> 00:05:20.720
there was a paper released yesterday, I believe,
00:05:20.860 --> 00:05:23.959
on describing a new Spinosaurus species, which
00:05:23.959 --> 00:05:26.379
the name I forget, but it's quite a similar species
00:05:26.379 --> 00:05:29.180
to Spinosaurus aegypticus, which you would have
00:05:29.180 --> 00:05:31.420
discussed, but it has this really cool head crest
00:05:31.420 --> 00:05:34.699
on the top of its head, probably for display
00:05:34.699 --> 00:05:38.500
purposes. It's a very interesting new discovery.
00:05:38.680 --> 00:05:41.519
So that's something your listeners might be...
00:05:42.009 --> 00:05:44.290
be interested to hear yeah to answer your question
00:05:44.290 --> 00:05:47.850
i think you know when i was a kid i thought dinosaurs
00:05:47.850 --> 00:05:51.430
were big and impressive and cool and as an adult
00:05:51.430 --> 00:05:54.730
uh i i actually think what makes them so interesting
00:05:54.730 --> 00:05:57.709
is the kind of mystery involved in understanding
00:05:57.709 --> 00:05:59.990
them because we you know we can't go back in
00:05:59.990 --> 00:06:02.509
time to see what they look like we we can't clone
00:06:02.509 --> 00:06:04.990
them and bring them back to life like uh in the
00:06:04.990 --> 00:06:07.069
movies so it's instead there's been these kind
00:06:07.069 --> 00:06:09.329
of centuries of detective work where you're kind
00:06:09.329 --> 00:06:11.990
of We're only just starting to scratch the surface
00:06:11.990 --> 00:06:14.209
of what they actually looked like and behaved
00:06:14.209 --> 00:06:17.129
like. And I think there are a lot of parallels
00:06:17.129 --> 00:06:21.189
between that and what we do at CGI. You know,
00:06:21.230 --> 00:06:24.410
we're doing geological and geophysical characterization.
00:06:24.529 --> 00:06:26.990
We're building methods that uncover the mysteries
00:06:26.990 --> 00:06:29.949
of the earth in the same way that paleontology
00:06:29.949 --> 00:06:32.589
uncovers the mysteries of ancient life. And I
00:06:32.589 --> 00:06:35.769
think both kind of, they require a certain imagination
00:06:35.769 --> 00:06:40.069
combined with evidence, like real evidence. So
00:06:40.069 --> 00:06:43.009
what you're saying is that your job today kind
00:06:43.009 --> 00:06:45.750
of marries some of the great things that you've
00:06:45.750 --> 00:06:48.149
always wanted to study, but provides a practical
00:06:48.149 --> 00:06:50.290
aspect to it moving forward, kind of like the
00:06:50.290 --> 00:06:52.550
industry and the like. Yeah, yeah, exactly. I
00:06:52.550 --> 00:06:55.110
think there's a common thread of, you know, I
00:06:55.110 --> 00:06:56.970
think all scientists have the kind of thirst
00:06:56.970 --> 00:07:00.470
for discovery in a way. Yeah, and that's like
00:07:00.470 --> 00:07:03.050
a perfect segue to the next question, because
00:07:03.050 --> 00:07:05.209
before we... kind of pivot towards the technical
00:07:05.209 --> 00:07:08.670
part the process of conducting science is just
00:07:08.670 --> 00:07:10.610
as important as the results sometimes and that's
00:07:10.610 --> 00:07:13.170
what we're talking about so here on whimsical
00:07:13.170 --> 00:07:15.029
wavelengths we've covered papers that have come
00:07:15.029 --> 00:07:17.850
from universities and museums mostly as far as
00:07:17.850 --> 00:07:20.670
like a direct paper from industry like we've
00:07:20.670 --> 00:07:22.689
had some industry types on but they didn't bring
00:07:22.689 --> 00:07:25.790
a paper they wrote for industry it was usually
00:07:25.790 --> 00:07:29.050
from it was either a topic or from their phd
00:07:29.050 --> 00:07:31.389
thesis or something like that and that's not
00:07:31.389 --> 00:07:34.850
a problem but This is pretty sure the first one
00:07:34.850 --> 00:07:37.410
directly from industry. Peer review is the same,
00:07:37.449 --> 00:07:39.769
though. But there are some differences. For us,
00:07:39.790 --> 00:07:42.490
because I'm also employed at Computational Geosciences,
00:07:42.490 --> 00:07:44.290
what's the biggest difference between industry
00:07:44.290 --> 00:07:48.110
science and what you would do just strictly for
00:07:48.110 --> 00:07:50.629
academics? Yeah, I've thought about this a lot,
00:07:50.670 --> 00:07:53.170
actually, because I obviously had a brief period
00:07:53.170 --> 00:07:56.009
in academia and published some papers and then
00:07:56.009 --> 00:07:58.379
took some... time off from the research side
00:07:58.379 --> 00:08:00.660
and now i'm kind of back in the area of publishing
00:08:00.660 --> 00:08:04.100
again i think the biggest difference is kind
00:08:04.100 --> 00:08:06.420
of the the consequence of actually being wrong
00:08:06.420 --> 00:08:09.480
with your results i think in academia if it doesn't
00:08:09.480 --> 00:08:12.920
uh quite work then you you might get your null
00:08:12.920 --> 00:08:15.079
results or whatever but and have to rethink your
00:08:15.079 --> 00:08:18.439
your project but um you know life goes on whereas
00:08:18.439 --> 00:08:21.439
uh if we if we tell a client to dig somewhere
00:08:21.439 --> 00:08:23.420
and it there's nothing there then that's a very
00:08:23.870 --> 00:08:27.370
expensive mistake for a client. So you're balancing
00:08:27.370 --> 00:08:31.850
this scientific discovery process using the science
00:08:31.850 --> 00:08:34.629
to find things with the kind of risk management
00:08:34.629 --> 00:08:37.210
side of things, like generating models, which
00:08:37.210 --> 00:08:39.830
are maybe slightly more cautious in terms of
00:08:39.830 --> 00:08:43.190
their predictive capabilities than otherwise.
00:08:43.490 --> 00:08:46.850
So I think it's the risk management side of things
00:08:46.850 --> 00:08:48.789
that's a bit different. Yeah. One of the other
00:08:48.789 --> 00:08:51.500
things I want to kind of poke at here is Peer
00:08:51.500 --> 00:08:53.539
review is the same, so once something gets through
00:08:53.539 --> 00:08:55.500
peer review, it doesn't really matter where it
00:08:55.500 --> 00:08:57.759
came from too, too much. Again, it's always one
00:08:57.759 --> 00:09:00.200
paper, so if it's industry, you should probably
00:09:00.200 --> 00:09:02.500
see it in an academic journal from an academic
00:09:02.500 --> 00:09:04.580
point of view eventually. You can't always just
00:09:04.580 --> 00:09:07.039
do one, but there's also a different emphasis
00:09:07.039 --> 00:09:09.419
on publishing in industry, isn't there? Like,
00:09:09.419 --> 00:09:12.379
you're not just in the pursuit of knowledge,
00:09:12.480 --> 00:09:16.059
you only publish kind of what you want to, right?
00:09:16.669 --> 00:09:18.669
Yeah, there's always like a financial incentive
00:09:18.669 --> 00:09:21.669
to it, I feel like. And in fact, a lot of people
00:09:21.669 --> 00:09:24.470
don't publish, right? Like a lot of companies
00:09:24.470 --> 00:09:26.990
will keep their trade secrets because that's
00:09:26.990 --> 00:09:30.850
their moneymaker. And yeah, I think I disagree
00:09:30.850 --> 00:09:33.610
with that. I think by publishing and discussing
00:09:33.610 --> 00:09:35.950
what we do at conferences and things, I think
00:09:35.950 --> 00:09:39.490
we're showing people that what we do is solid
00:09:39.490 --> 00:09:42.690
and backed by science and open to public scrutiny.
00:09:43.049 --> 00:09:46.299
I think that kind of... transparency builds a
00:09:46.299 --> 00:09:50.419
certain trust in what we do and I think if we
00:09:50.419 --> 00:09:54.960
can prove what we do in areas like this paper
00:09:54.960 --> 00:09:58.519
we're doing, we're studying some areas in Australia,
00:09:58.740 --> 00:10:00.460
I think if we can prove that it works there then
00:10:00.460 --> 00:10:04.320
people trust it when we suggest using it in other
00:10:04.320 --> 00:10:07.639
areas where people haven't looked before. All
00:10:07.639 --> 00:10:09.559
right, so the paper we're actually about to dive
00:10:09.559 --> 00:10:11.879
into with this Australia data is formulating
00:10:11.879 --> 00:10:15.899
gold prospectivity mapping as a constrained learning
00:10:15.899 --> 00:10:18.059
problem. It was submitted and accepted to the
00:10:18.059 --> 00:10:22.019
Society of Exploration Geophysics for a meeting
00:10:22.019 --> 00:10:24.419
last year for the Applied Geoscience and Energy.
00:10:25.799 --> 00:10:28.899
Is there anything that you need to add to from
00:10:28.899 --> 00:10:32.159
the preamble that was part of the intro, or what
00:10:32.159 --> 00:10:35.529
problem is this trying to address? By the way,
00:10:35.529 --> 00:10:38.110
quick update. It should be published this month.
00:10:38.250 --> 00:10:39.549
Again, I'm not sure when you're releasing this,
00:10:39.629 --> 00:10:41.730
but it should be published in February in the
00:10:41.730 --> 00:10:46.049
SEG library. So it should be accessible by the
00:10:46.049 --> 00:10:47.929
time this comes out, I would imagine. But yes,
00:10:48.070 --> 00:10:51.490
so no, I think you did a good job. I think the
00:10:51.490 --> 00:10:55.490
motivation really is how complicated mineral
00:10:55.490 --> 00:10:59.250
systems are. They can occur on these tiny little
00:10:59.250 --> 00:11:01.309
centimeter length scales, and they're formed
00:11:01.309 --> 00:11:03.730
by incredibly complex and poorly understood.
00:11:04.279 --> 00:11:08.519
geological causation. And traditional exploration,
00:11:08.659 --> 00:11:10.759
I think, is very good at filling in the gaps
00:11:10.759 --> 00:11:14.299
between your knowns. But it's the extrapolating
00:11:14.299 --> 00:11:17.139
which is the difficult part and the thing that
00:11:17.139 --> 00:11:22.039
AI prospectivity is attempting to approach. What
00:11:22.039 --> 00:11:24.360
we're trying to do is to build a model which
00:11:24.360 --> 00:11:27.620
can do that, can extrapolate from a well -known
00:11:27.620 --> 00:11:29.500
area into an area that it doesn't know about
00:11:29.500 --> 00:11:32.529
at all, whilst also... Managing that risk, again,
00:11:32.590 --> 00:11:34.809
it comes back to managing risk, managing the
00:11:34.809 --> 00:11:37.690
risk of the model hallucinating or getting confused
00:11:37.690 --> 00:11:40.529
when we move away from the safe areas, I guess.
00:11:41.090 --> 00:11:43.730
Okay, so let's expand some of those ideas out
00:11:43.730 --> 00:11:46.250
right here, I guess. Prospectivity and machine
00:11:46.250 --> 00:11:48.509
learning is essentially, it's the machine learning
00:11:48.509 --> 00:11:50.149
side of it that you're talking about, the networking.
00:11:50.450 --> 00:11:53.409
So the average person listening to this has heard
00:11:53.409 --> 00:11:56.850
of large language models like ChatGPT and that
00:11:56.850 --> 00:11:59.009
they're trained by scraping art and music and
00:11:59.009 --> 00:12:02.019
writing from the internet. in an effort to learn
00:12:02.019 --> 00:12:05.200
to respond more human -like and give information
00:12:05.200 --> 00:12:08.480
and answers. So leaving the pitfalls and issues
00:12:08.480 --> 00:12:12.220
with that process aside for now, so how does
00:12:12.220 --> 00:12:14.940
that work in principle here? Because that's what
00:12:14.940 --> 00:12:16.980
you're talking about, being able to extrapolate
00:12:16.980 --> 00:12:19.720
from a well -known area to a less known area.
00:12:19.799 --> 00:12:24.019
I'm assuming you're having to train on data very
00:12:24.019 --> 00:12:28.600
similarly to LLMs. How is it different and how
00:12:28.600 --> 00:12:31.899
is that kind of all taking place? Yeah, the concept
00:12:31.899 --> 00:12:34.840
is very similar. So as you say, ChatGPT has learned
00:12:34.840 --> 00:12:39.179
how humans communicate via its training data,
00:12:39.340 --> 00:12:41.720
which is the internet, and it's able to extrapolate
00:12:41.720 --> 00:12:44.799
its responses to things it hasn't heard before.
00:12:45.399 --> 00:12:49.200
And what we're doing is we're building an algorithm
00:12:49.200 --> 00:12:52.259
which has, in principle, if we've done our job,
00:12:52.360 --> 00:12:55.759
it's learned how mineral systems form via the...
00:12:56.059 --> 00:12:59.000
geological training data, and it's extrapolating
00:12:59.000 --> 00:13:02.299
those, I guess, like formation mechanisms to
00:13:02.299 --> 00:13:04.820
areas that it hasn't seen before in the same
00:13:04.820 --> 00:13:07.700
way ChatGPT has. So that's the similarity. I
00:13:07.700 --> 00:13:10.620
think the difference is the amount of data. I
00:13:10.620 --> 00:13:12.700
think ChatGPT obviously had the whole internet
00:13:12.700 --> 00:13:15.159
to scrape from. We don't quite have that privilege,
00:13:15.320 --> 00:13:19.500
so ChatGPT achieves. like uh insanely creepy
00:13:19.500 --> 00:13:21.879
level of human i don't know if you ever like
00:13:21.879 --> 00:13:23.460
speak to it you ever do like a back and forth
00:13:23.460 --> 00:13:25.600
conversation with it it's it's it's creepy how
00:13:25.600 --> 00:13:29.100
yeah and full disclosure i often give it podcast
00:13:29.100 --> 00:13:32.120
stuff and ask it to like critique my stuff and
00:13:32.120 --> 00:13:34.059
sometimes the stuff it gives me back you're like
00:13:34.059 --> 00:13:39.110
oh yeah Yeah, we don't quite have that level
00:13:39.110 --> 00:13:40.870
of training data. So again, it's back to that
00:13:40.870 --> 00:13:43.649
risk management thing. It's about how to make
00:13:43.649 --> 00:13:46.889
it work with the limited data that we have in
00:13:46.889 --> 00:13:49.570
the geological world. Because, you know, drilling
00:13:49.570 --> 00:13:52.389
is expensive in geological exploration. Geochemical
00:13:52.389 --> 00:13:56.149
sampling is often quite sparse. So it's about
00:13:56.149 --> 00:13:58.950
all those kind of limitations. Yeah, so maybe
00:13:58.950 --> 00:14:01.110
we'll take a quick aside there. Do you happen
00:14:01.110 --> 00:14:02.830
to know, roughly speaking, how much a drill hole
00:14:02.830 --> 00:14:06.190
costs? I actually have no idea, no. Oh, okay.
00:14:06.269 --> 00:14:09.029
Well, I guess for the listener, it's usually
00:14:09.029 --> 00:14:10.750
in the many hundreds of thousands of dollars,
00:14:10.850 --> 00:14:12.909
depending on how many you're going to get done.
00:14:13.009 --> 00:14:14.529
If you're just going to go do one drill hole,
00:14:14.629 --> 00:14:16.450
it's like a million dollars because you have
00:14:16.450 --> 00:14:19.830
to pay to get everybody on site. But then it's
00:14:19.830 --> 00:14:22.289
usually somewhere between $200 ,000 to $300 ,000
00:14:22.289 --> 00:14:24.820
per drill hole. I guess the depth as well depends
00:14:24.820 --> 00:14:26.500
on the depth. It depends on the depth. It depends
00:14:26.500 --> 00:14:27.860
on a lot of things. So if you're going to go
00:14:27.860 --> 00:14:29.840
really deep, then it's almost a million per drill
00:14:29.840 --> 00:14:32.100
hole. Or if you're going to do geothermal because
00:14:32.100 --> 00:14:34.340
of the size of the drill hole, it has to be,
00:14:34.379 --> 00:14:37.360
it'll be more expensive. But the point is being
00:14:37.360 --> 00:14:40.039
able to de -risk those drill holes is really
00:14:40.039 --> 00:14:44.480
important for the industry. So as you were saying,
00:14:44.580 --> 00:14:47.940
you don't have the same amounts of data as a
00:14:47.940 --> 00:14:49.840
lot of these LLMs because you're dealing with
00:14:49.840 --> 00:14:52.879
an area which we only can, We only have what
00:14:52.879 --> 00:14:55.360
information we have is what we can measure. So
00:14:55.360 --> 00:14:59.159
the paper uses data that is either already public
00:14:59.159 --> 00:15:01.360
or you've received some kind of special permission
00:15:01.360 --> 00:15:04.200
to be able to publish it, right? So what is that
00:15:04.200 --> 00:15:06.840
data that you're using to train in this paper,
00:15:06.860 --> 00:15:10.059
which is on Australia, and build that geological
00:15:10.059 --> 00:15:13.679
equivalent to ChatGPT? Yeah, so we're focusing
00:15:13.679 --> 00:15:16.600
on Western Australia and Australia is very good
00:15:16.600 --> 00:15:20.899
at releasing their data for public use. There's
00:15:20.899 --> 00:15:25.259
a website called Geoscience Australia that merges
00:15:25.259 --> 00:15:27.919
and makes available a lot of exploration data
00:15:27.919 --> 00:15:30.299
and they have essentially geophysics for all
00:15:30.299 --> 00:15:33.799
of Western Australia, which is absolutely ideal
00:15:33.799 --> 00:15:39.000
for a test project like this. So we went on Geoscience
00:15:39.000 --> 00:15:41.240
Australia, we assembled what we call a data cube,
00:15:41.379 --> 00:15:44.039
which is just a fancy name for a big folder of
00:15:44.039 --> 00:15:46.919
data. But we assembled, I think, something like
00:15:46.919 --> 00:15:49.740
15 different data layers of public data from
00:15:49.740 --> 00:15:52.759
the Yilgarn Craton, which is an area of Western
00:15:52.759 --> 00:15:55.200
Australia. And that includes potential fields
00:15:55.200 --> 00:15:58.379
data, which is your... magnetics and gravity
00:15:58.379 --> 00:16:02.179
data, the images into the subsurface. We have
00:16:02.179 --> 00:16:05.120
radiometrics, we have satellite imagery, and
00:16:05.120 --> 00:16:08.200
then we have the labels, which are geochemical
00:16:08.200 --> 00:16:10.940
samples and potentially the most important part
00:16:10.940 --> 00:16:13.840
of the whole puzzle. And the idea is to build
00:16:13.840 --> 00:16:17.639
up an analog to doing this on a more kind of
00:16:17.639 --> 00:16:21.240
brownfield level, I guess, like the kind of data
00:16:21.240 --> 00:16:23.700
that a client might potentially have at a kind
00:16:23.700 --> 00:16:26.679
of exploration target. Okay. only because I know
00:16:26.679 --> 00:16:28.940
the words, and I'm going to try to expand them
00:16:28.940 --> 00:16:30.980
out here a little bit. When we deal with, I guess,
00:16:31.059 --> 00:16:33.460
potential fields, we're talking about someone's
00:16:33.460 --> 00:16:35.940
gone around and measured the Earth's magnetic
00:16:35.940 --> 00:16:38.639
field or gravitational field over a wide area,
00:16:38.740 --> 00:16:41.320
and you're taking that map and derivatives of
00:16:41.320 --> 00:16:44.679
that map as some of your data, right? Yeah, exactly.
00:16:45.450 --> 00:16:48.169
And for those that are not in the industry, brownfields
00:16:48.169 --> 00:16:51.830
means like you've had enough time to collect
00:16:51.830 --> 00:16:54.149
a lot of different data. You know that there's
00:16:54.149 --> 00:16:56.009
something under the ground and you're moving
00:16:56.009 --> 00:16:59.129
towards hopefully opening a mine. In a lot of
00:16:59.129 --> 00:17:01.429
cases when we're looking for minerals, we call
00:17:01.429 --> 00:17:03.809
it greenfields when we don't have a lot of data
00:17:03.809 --> 00:17:06.410
and you've maybe got one or two drill holes kind
00:17:06.410 --> 00:17:09.089
of idea. still trying to suss out whether or
00:17:09.089 --> 00:17:11.029
not there's anything of economic worth underneath
00:17:11.029 --> 00:17:12.970
the ground. Whereas brownfields, you're starting
00:17:12.970 --> 00:17:15.589
to move to words of mine. That doesn't mean you'll
00:17:15.589 --> 00:17:16.990
get there, but that's where you're moving to.
00:17:17.049 --> 00:17:19.930
So you have a lot of data, or at least a lot
00:17:19.930 --> 00:17:24.390
of data in the sense of geology and data sets
00:17:24.390 --> 00:17:27.369
that I guess you can use for prospectivity. That
00:17:27.369 --> 00:17:28.630
summed it up pretty good. Did I miss anything?
00:17:29.210 --> 00:17:33.109
No. Yeah, perfect. And sorry, I'm used to presenting
00:17:33.109 --> 00:17:35.230
this at conferences and things where the terminology
00:17:35.230 --> 00:17:37.559
is more known. Yeah, but you're also talking
00:17:37.559 --> 00:17:40.119
to me, a co -worker who is going to understand
00:17:40.119 --> 00:17:42.920
what you say kind of instantaneously instead
00:17:42.920 --> 00:17:45.720
of for everyone who's listening somewhere else.
00:17:45.859 --> 00:17:48.480
I mean, don't be afraid to get nerdy, my friend.
00:17:48.619 --> 00:17:51.299
I mean, we all love maps, or at least we do.
00:17:52.140 --> 00:17:54.400
So I guess the next part of the conversation
00:17:54.400 --> 00:17:56.339
comes back to the geochem that you were just
00:17:56.339 --> 00:17:58.759
saying. The idea of like, we've already talked
00:17:58.759 --> 00:18:01.259
about the different data layers. But the data
00:18:01.259 --> 00:18:04.240
layers isn't really what you're training on,
00:18:04.380 --> 00:18:06.819
right? That's just giving the information for
00:18:06.819 --> 00:18:11.059
it to make pattern recognition with geologic
00:18:11.059 --> 00:18:14.980
layers or geologic targets versus what LLMs do
00:18:14.980 --> 00:18:18.720
with words, right? So you need to give it something,
00:18:18.920 --> 00:18:22.460
and you call that the labels. Yeah, yeah. So
00:18:22.460 --> 00:18:25.599
in machine learning, you classify your data into
00:18:25.599 --> 00:18:30.569
features and labels. So features are... the kind
00:18:30.569 --> 00:18:33.970
of wide area extensive geophysics in this case
00:18:33.970 --> 00:18:36.549
or remote sensing and labels are the actual ground
00:18:36.549 --> 00:18:39.950
truth so the the concept of machine learning
00:18:39.950 --> 00:18:42.430
is that you're building this kind of causative
00:18:42.430 --> 00:18:45.789
link between your labels and your features so
00:18:45.789 --> 00:18:49.009
the the features are the observable thing which
00:18:49.009 --> 00:18:51.910
explain the existence of the labels so in this
00:18:51.910 --> 00:18:54.170
case the the ground truth from from the drilling
00:18:54.170 --> 00:18:57.390
is the labels and the geophysics is the thing
00:18:57.390 --> 00:19:00.250
which is kind of explaining the the reason for
00:19:00.250 --> 00:19:03.589
the labels being what they are. Okay. Just out
00:19:03.589 --> 00:19:06.710
of curiosity, there must be some like pitfalls
00:19:06.710 --> 00:19:09.009
to this type of approach as well, right? Like
00:19:09.009 --> 00:19:11.670
in my own head, I come up with the simple thing
00:19:11.670 --> 00:19:14.009
is that if you're using ground geochemistry,
00:19:14.089 --> 00:19:17.910
something could have changed for the lack of
00:19:17.910 --> 00:19:20.250
a, obviously we wouldn't use this necessarily,
00:19:20.329 --> 00:19:24.710
but glaciers or rivers or flash floods can move
00:19:24.710 --> 00:19:29.250
things downstream and modify what we see. Is
00:19:29.250 --> 00:19:32.009
that a concern here? Is it things that modify
00:19:32.009 --> 00:19:34.569
either humans or natural events that modify the
00:19:34.569 --> 00:19:37.369
subsurface? Is it a case of garbage in, garbage
00:19:37.369 --> 00:19:41.690
out? Yeah, absolutely. That's the big caveat
00:19:41.690 --> 00:19:43.970
of all AI is the garbage in, garbage out thing.
00:19:44.450 --> 00:19:47.430
And geochemical sampling is especially prone
00:19:47.430 --> 00:19:50.509
to that human or instrument error or transported
00:19:50.509 --> 00:19:54.470
samples, like you were saying. And when you're
00:19:54.470 --> 00:19:56.910
building that causative link I mentioned, then...
00:19:57.279 --> 00:19:58.960
that can obviously throw off your entire model.
00:19:59.180 --> 00:20:01.859
So yeah, garbage in, garbage out for sure. I
00:20:01.859 --> 00:20:03.339
mean, again, it comes back to that risk management
00:20:03.339 --> 00:20:06.279
thing, building these for real world applications.
00:20:06.380 --> 00:20:08.839
And that kind of prompted our approach that we
00:20:08.839 --> 00:20:11.180
present in this paper. Actually, there's a nice
00:20:11.180 --> 00:20:13.200
lead into that because we don't actually care
00:20:13.200 --> 00:20:16.460
in our approach about the exact granular noise
00:20:16.460 --> 00:20:19.400
of a geochemical sample. We're kind of more looking
00:20:19.400 --> 00:20:22.000
at the bigger picture, the proximity to the kind
00:20:22.000 --> 00:20:26.410
of high grade hits rather than the misses. So
00:20:26.410 --> 00:20:28.289
it's more resilient to that garbage because we're
00:20:28.289 --> 00:20:30.890
kind of building this sort of probability model
00:20:30.890 --> 00:20:32.750
based on the bigger picture rather than the kind
00:20:32.750 --> 00:20:35.970
of individual samples. OK, and that's what is
00:20:35.970 --> 00:20:38.569
new and novel with the paper. Is that specifically?
00:20:39.289 --> 00:20:41.970
Yes, exactly. I think we've kind of skipped over
00:20:41.970 --> 00:20:44.250
one part of this. So we've just kind of I've
00:20:44.250 --> 00:20:45.910
kind of waved a magic wand and said it's machine
00:20:45.910 --> 00:20:49.029
learning. We haven't really described what machine
00:20:49.029 --> 00:20:52.210
learning is. Yeah, well, machine learning is
00:20:52.210 --> 00:20:55.859
is kind of. building that kind of causative link
00:20:55.859 --> 00:20:57.940
that I mentioned between some kind of feature
00:20:57.940 --> 00:21:00.059
and some kind of label, and building a model
00:21:00.059 --> 00:21:02.319
that understands that and is then able to extrapolate
00:21:02.319 --> 00:21:06.099
it to kind of unknown areas where we don't necessarily
00:21:06.099 --> 00:21:09.420
have that understanding. Yeah, I was going to
00:21:09.420 --> 00:21:11.599
give a bit more into the details, really get
00:21:11.599 --> 00:21:14.259
nerdy here. Neural networks is more where I'm
00:21:14.259 --> 00:21:16.019
going because there's different types, right?
00:21:16.160 --> 00:21:19.220
Not all machine learning is the same, built the
00:21:19.220 --> 00:21:21.930
same way. yeah for sure so we're using a neural
00:21:21.930 --> 00:21:24.049
network in this paper and a neural network is
00:21:24.049 --> 00:21:26.450
based on the architecture of a human brain essentially
00:21:26.450 --> 00:21:30.609
so um it's has various kind of neural nodes and
00:21:30.609 --> 00:21:32.710
and you know connections between those nodes
00:21:32.710 --> 00:21:36.130
and is assembling a uh kind of unquantifiable
00:21:36.130 --> 00:21:39.930
knowledge of of those relationships uh in a in
00:21:39.930 --> 00:21:43.230
a different way to the way that like a regression
00:21:43.230 --> 00:21:45.150
is a machine learning method but that's learning
00:21:45.150 --> 00:21:47.309
a very like simple linear relationship between
00:21:47.309 --> 00:21:52.309
things this is a very complex, non -linear, like
00:21:52.309 --> 00:21:56.109
very hard to even sort of picture in your head
00:21:56.109 --> 00:21:58.210
really what a neural network looks like. It's
00:21:58.210 --> 00:22:02.650
kind of beyond visualization. And that kind of
00:22:02.650 --> 00:22:06.569
gets to how advanced these methods are. Complete
00:22:06.569 --> 00:22:09.210
aside, you said it's a lot like the brain. Do
00:22:09.210 --> 00:22:13.009
the two fields actually kind of intersect there
00:22:13.009 --> 00:22:16.269
a little bit? What do you mean? Are we programming
00:22:16.269 --> 00:22:18.690
human beings here? Are we actually taking what
00:22:18.690 --> 00:22:21.630
we understand of a brain and actually utilizing
00:22:21.630 --> 00:22:23.490
that information to create a neural network?
00:22:23.849 --> 00:22:26.509
Yeah, in a way, you can kind of look at it in
00:22:26.509 --> 00:22:29.410
that we're sort of creating an artificial geologist
00:22:29.410 --> 00:22:33.069
in a lot of ways. A geologist that is able to
00:22:33.069 --> 00:22:36.589
look at data and kind of use their own understanding
00:22:36.589 --> 00:22:38.950
of the connections between the data and then
00:22:38.950 --> 00:22:41.829
make predictions based on what it sees. Obviously,
00:22:41.829 --> 00:22:43.740
it's a bit more... complicated than that. And
00:22:43.740 --> 00:22:46.220
we don't want to replace our geologist friends.
00:22:46.420 --> 00:22:48.640
We're always going to need their inputs. And
00:22:48.640 --> 00:22:52.220
these things are being built as tools for geologists.
00:22:52.420 --> 00:22:56.640
But it is essentially a brain that we're creating
00:22:56.640 --> 00:22:59.859
here. Yeah, I guess it takes a geologist to understand
00:22:59.859 --> 00:23:02.880
the output. Yeah. You're never going to replace
00:23:02.880 --> 00:23:05.900
the geologist because they have to feed it. And
00:23:05.900 --> 00:23:07.839
then they have to understand the output. Yeah,
00:23:07.880 --> 00:23:10.470
absolutely. I guess that gets to the next part
00:23:10.470 --> 00:23:12.009
of the question. Like AI and machine learning,
00:23:12.089 --> 00:23:14.470
it's revolutionizing some aspects. In some cases,
00:23:14.529 --> 00:23:19.549
it's more, instead of revolutionizing something,
00:23:19.710 --> 00:23:24.369
it's like methods of marketing. I guess this
00:23:24.369 --> 00:23:26.329
happens every time there's a new technological
00:23:26.329 --> 00:23:29.329
disruptor that's out there. But this is peer
00:23:29.329 --> 00:23:31.630
-reviewed. It's solid science backing with real
00:23:31.630 --> 00:23:34.329
-world results. What's the strategy to break
00:23:34.329 --> 00:23:36.450
through the noise around AI and deliver real
00:23:36.450 --> 00:23:38.589
results and get people to believe in it? I mean,
00:23:38.609 --> 00:23:40.490
it's not really to believe in it, but to understand
00:23:40.490 --> 00:23:42.730
the results and how they can use it. There's
00:23:42.730 --> 00:23:46.930
just so much out there right now. I mean, I wish
00:23:46.930 --> 00:23:50.690
I knew the best strategy, Jeff. I think our strategy
00:23:50.690 --> 00:23:54.549
is to just kind of be as honest, like scientifically
00:23:54.549 --> 00:23:57.789
honest as possible about the capabilities of
00:23:57.789 --> 00:23:59.910
this and also the limitations. I think a lot
00:23:59.910 --> 00:24:03.049
of those kind of hype companies don't. discuss
00:24:03.049 --> 00:24:05.670
the limitations and that's what makes it snake
00:24:05.670 --> 00:24:08.769
oil in my belief. I think what we're doing by
00:24:08.769 --> 00:24:11.430
regularly publishing and talking about what we
00:24:11.430 --> 00:24:16.289
do is building that trust and acknowledging the
00:24:16.289 --> 00:24:20.490
inherent faults of all AI kind of allows us to
00:24:20.490 --> 00:24:22.650
actually work around those faults in a smart
00:24:22.650 --> 00:24:25.690
way by integrating as much input and knowledge
00:24:25.690 --> 00:24:29.309
from geologists and field experts to kind of
00:24:29.309 --> 00:24:32.069
build that tool rather than the sort of vague
00:24:32.069 --> 00:24:35.690
black box that I think has a lot of pitfalls
00:24:35.690 --> 00:24:38.609
in our industry. Okay, so let's pivot back to
00:24:38.609 --> 00:24:41.349
the results. So this is an evolving tool to find
00:24:41.349 --> 00:24:45.990
mineral deposits, but what did you find by going
00:24:45.990 --> 00:24:48.369
through this study? Was there anything new and
00:24:48.369 --> 00:24:51.529
surprising that we can discuss here? Yeah, I
00:24:51.529 --> 00:24:54.930
mean, there's nothing particularly revolutionary
00:24:54.930 --> 00:24:57.509
about what we did. It's more kind of how we framed
00:24:57.509 --> 00:25:02.079
the problem, I guess. In terms of where this
00:25:02.079 --> 00:25:04.019
will be going in future, I think the holy grail
00:25:04.019 --> 00:25:06.400
is to integrate that kind of 3D information.
00:25:06.599 --> 00:25:12.019
This is a very 2D problem. And I think we want
00:25:12.019 --> 00:25:14.380
to... There's one thing like predicting where
00:25:14.380 --> 00:25:17.380
to drill on a map, but actually like how deep
00:25:17.380 --> 00:25:19.960
to drill is kind of the holy grail in my opinion.
00:25:20.160 --> 00:25:24.319
So I think the kind of evolving steps are to
00:25:24.319 --> 00:25:27.299
look into hybrid approaches that kind of combine
00:25:27.299 --> 00:25:31.309
the advantages of... neural networks and AI with
00:25:31.309 --> 00:25:37.009
traditional geophysical inversions and 3D interpolations
00:25:37.009 --> 00:25:40.369
as well. And I think we hope to be able to talk
00:25:40.369 --> 00:25:42.849
about that in the near future. Yeah, I guess
00:25:42.849 --> 00:25:44.390
we didn't even touch on the fact that this is
00:25:44.390 --> 00:25:47.109
really providing a 2D map, right? Like the idea
00:25:47.109 --> 00:25:50.369
is you're providing a heat map or something like
00:25:50.369 --> 00:25:53.349
a heat map that you get warmer colors when you're
00:25:53.349 --> 00:25:55.410
closer to what you're looking for. In this case
00:25:55.410 --> 00:25:59.289
was gold, right? Yes, exactly. Yeah. So I guess
00:25:59.289 --> 00:26:02.170
the obvious question is, why can't we do it in
00:26:02.170 --> 00:26:04.990
3D? Like, what's the limitation? I think that
00:26:04.990 --> 00:26:08.309
there is a computational limitation. I think
00:26:08.309 --> 00:26:12.410
it's also just the kind of how you frame the
00:26:12.410 --> 00:26:14.490
problem. I think this is a very unconstrained
00:26:14.490 --> 00:26:17.289
kind of solution that we've developed. I think
00:26:17.289 --> 00:26:21.720
a 3D solution because it's in so much more. resolution
00:26:21.720 --> 00:26:23.779
and there's so much more information, I think
00:26:23.779 --> 00:26:26.180
you have to introduce a lot of constraints to
00:26:26.180 --> 00:26:28.940
that in a positive way. I think that's a good
00:26:28.940 --> 00:26:31.819
thing to produce models which are consistent
00:26:31.819 --> 00:26:37.259
with your actual 3D information. You just have
00:26:37.259 --> 00:26:39.140
to parameterize that a little bit differently
00:26:39.140 --> 00:26:43.559
and it's not something which is unachievable
00:26:43.559 --> 00:26:47.650
and it is something that we are working on. So
00:26:47.650 --> 00:26:51.089
you would say it's more a function of not having
00:26:51.089 --> 00:26:54.890
3D data? Like we have some, but again, being
00:26:54.890 --> 00:26:58.289
a geophysicist and we've covered inversions before,
00:26:58.509 --> 00:27:02.809
inversions are non -unique. So you can use them,
00:27:02.890 --> 00:27:05.769
but you can't use them for your labels, right?
00:27:06.549 --> 00:27:12.890
Yeah, exactly. An inversion is a subjective interpretation
00:27:12.890 --> 00:27:17.859
of objective data. And that's where the non -uniqueness
00:27:17.859 --> 00:27:20.759
comes from. And there's an infinite amount of
00:27:20.759 --> 00:27:26.960
models which satisfy a given data set. And we
00:27:26.960 --> 00:27:29.839
do things in Inversion to try and reduce the
00:27:29.839 --> 00:27:32.380
amount of models which do satisfy that data set.
00:27:32.500 --> 00:27:36.059
But the more you can constrain and limit that
00:27:36.059 --> 00:27:41.400
process, the better. So I guess more to the frame,
00:27:41.539 --> 00:27:44.119
I'm familiar with geophysics and Inversion models.
00:27:44.809 --> 00:27:47.210
And I guess I'm familiar with geologic models,
00:27:47.369 --> 00:27:49.069
but generally speaking, there isn't that many
00:27:49.069 --> 00:27:52.109
geologic models out there to be had that aren't
00:27:52.109 --> 00:27:54.150
similar to inversion models, that they're just
00:27:54.150 --> 00:27:56.170
interpolated between. We don't have a lot of
00:27:56.170 --> 00:27:58.589
labels. So is part of the reason why we can't
00:27:58.589 --> 00:28:02.089
have these 3D prospectivity more to the fact
00:28:02.089 --> 00:28:04.509
that we just don't have 3D data? Like, how do
00:28:04.509 --> 00:28:07.930
you bring a DEM into depth or things that you
00:28:07.930 --> 00:28:09.950
collect on the Earth's surface into something
00:28:09.950 --> 00:28:12.839
you can actually... The only thing I can think
00:28:12.839 --> 00:28:16.079
of we have is the drill holes. Yeah, and drill
00:28:16.079 --> 00:28:18.500
holes are obviously incredibly valuable that
00:28:18.500 --> 00:28:21.500
they're expensive, but that's an objective observation
00:28:21.500 --> 00:28:24.619
of what the lithology actually looks like. And
00:28:24.619 --> 00:28:27.920
you often get geophysical information with that
00:28:27.920 --> 00:28:30.660
too. So I think the real value comes from combining
00:28:30.660 --> 00:28:33.779
that knowledge of lithology, knowledge of downhole
00:28:33.779 --> 00:28:35.980
geophysics, and then combining that with surface
00:28:35.980 --> 00:28:39.400
geophysics. and kind of building something that
00:28:39.400 --> 00:28:43.359
satisfies all of those at the same time. And
00:28:43.359 --> 00:28:45.099
it's a slightly different problem to the one
00:28:45.099 --> 00:28:47.000
we tackled in this paper. This paper was more
00:28:47.000 --> 00:28:51.140
about a kind of rapid 2D kind of de -risking
00:28:51.140 --> 00:28:55.240
product. It's not supposed to be objective or
00:28:55.240 --> 00:28:59.059
an accurate model of the subsurface. It's supposed
00:28:59.059 --> 00:29:02.619
to be something for simple de -risking and for
00:29:02.619 --> 00:29:06.609
exploration targeting. I think what... The next
00:29:06.609 --> 00:29:10.750
step is this kind of 3D modeling, which builds
00:29:10.750 --> 00:29:13.509
a slightly more targetable model in terms of
00:29:13.509 --> 00:29:16.410
drilling. Yeah, and actually, there's a couple
00:29:16.410 --> 00:29:18.710
of ways I've also seen this used. You're probably
00:29:18.710 --> 00:29:21.349
aware of them as well. Like in British Columbia,
00:29:21.470 --> 00:29:25.589
they sent out a request for proposals from companies
00:29:25.589 --> 00:29:27.910
such as ours. We didn't take them up on this
00:29:27.910 --> 00:29:30.529
to figure out where places were perspective,
00:29:30.769 --> 00:29:32.829
so that way they could make decisions based on...
00:29:33.309 --> 00:29:36.730
parks and i guess first nations rights and stuff
00:29:36.730 --> 00:29:39.309
like that and where to start that initial conversations
00:29:39.309 --> 00:29:42.269
are there other examples where it's being used
00:29:42.269 --> 00:29:44.750
to inform policy instead of directly trying to
00:29:44.750 --> 00:29:47.410
find the next deposit no i mostly people are
00:29:47.410 --> 00:29:50.109
quite distrustful of this i feel like um it's
00:29:50.109 --> 00:29:52.990
that there's a few like national efforts to to
00:29:52.990 --> 00:29:56.089
do mineral prospectivity generally they're quite
00:29:56.089 --> 00:29:58.549
simplistic and i don't quite know the use cases
00:29:58.549 --> 00:30:02.440
for them i think I think generally the main driver
00:30:02.440 --> 00:30:05.819
is coming from industry, from mines and from
00:30:05.819 --> 00:30:08.480
that kind of brownfield, which means on mine
00:30:08.480 --> 00:30:11.859
site kind of targeting. So you're saying perhaps
00:30:11.859 --> 00:30:16.700
most of the real world case examples are locked
00:30:16.700 --> 00:30:19.140
behind data vaults that never get published.
00:30:19.819 --> 00:30:22.579
Exactly, yeah. What about other things that this
00:30:22.579 --> 00:30:25.259
same technology can be used for? Like, they can't
00:30:25.259 --> 00:30:27.140
be just for finding minerals. There's got to
00:30:27.140 --> 00:30:32.380
be water resources. identifying things on the
00:30:32.380 --> 00:30:35.720
Earth's surface? Yeah, we never reinvented the
00:30:35.720 --> 00:30:38.660
wheel with this technique. We produced a novel
00:30:38.660 --> 00:30:40.759
approach to it, but the actual technique is very
00:30:40.759 --> 00:30:43.839
grounded in decades of research. So this is a
00:30:43.839 --> 00:30:46.720
common thing in neuroscience and in medical sciences.
00:30:47.440 --> 00:30:50.579
It's the same technology that you would use to
00:30:50.579 --> 00:30:53.240
detect brain tumors, for example. You would train
00:30:53.240 --> 00:30:55.740
the model to detect the... the causative link
00:30:55.740 --> 00:31:00.559
between MRI images and brain tumors. And we actually
00:31:00.559 --> 00:31:04.339
started this study with a proof of concept using
00:31:04.339 --> 00:31:07.140
this malarial prevention data set, which was
00:31:07.140 --> 00:31:10.140
satellite imagery, or aerial imagery, sorry,
00:31:10.220 --> 00:31:15.240
of water tanks in Central America for the purpose
00:31:15.240 --> 00:31:18.579
of mapping the locations where mosquitoes might
00:31:18.579 --> 00:31:21.710
breed. And that was kind of our proof of concept
00:31:21.710 --> 00:31:25.069
was identifying water tanks in satellite, sorry,
00:31:25.690 --> 00:31:30.490
aerial data. This is a very kind of unknown technique.
00:31:30.549 --> 00:31:34.130
And the novel thing is just kind of how we framed
00:31:34.130 --> 00:31:38.390
that problem. So we just went from mineral prospectivity
00:31:38.390 --> 00:31:42.930
to medical imaging to finding where mosquitoes
00:31:42.930 --> 00:31:45.529
breed to help control malaria all in the same
00:31:45.529 --> 00:31:49.069
sentence. Yeah. Yeah, a bit of a roller coaster.
00:31:49.509 --> 00:31:52.829
No, but I mean, that just shows like how wide
00:31:52.829 --> 00:31:56.329
this data or this technique or general science
00:31:56.329 --> 00:31:58.509
behind the technique can be used for, right?
00:31:58.609 --> 00:32:00.650
Like this is going to have real world implications.
00:32:01.230 --> 00:32:05.069
Yeah, absolutely. And I think it will change
00:32:05.069 --> 00:32:08.309
the world of mineral exploration eventually.
00:32:08.490 --> 00:32:10.390
I don't know if it's quite there yet, but it's
00:32:10.390 --> 00:32:14.029
something that will only grow over time. So before
00:32:14.029 --> 00:32:17.119
we go to the traditional last question. All things
00:32:17.119 --> 00:32:18.359
being equal, is there anything else you want
00:32:18.359 --> 00:32:20.880
to share science -wise? A future trip, a paper,
00:32:21.019 --> 00:32:23.400
a documentary suggestion? I don't know, is there
00:32:23.400 --> 00:32:26.660
a new Spinosaurus documentary coming out? I don't
00:32:26.660 --> 00:32:29.180
know, that would be cool. I mentioned the Spinosaurus
00:32:29.180 --> 00:32:31.220
paper earlier. I recommend checking that out.
00:32:31.680 --> 00:32:35.599
My latest extracurricular obsession has been...
00:32:35.839 --> 00:32:39.160
something called bioacoustics which is uh studying
00:32:39.160 --> 00:32:42.180
the the vocalizations of marine life because
00:32:42.180 --> 00:32:44.519
i'm really into whale watching and so i've been
00:32:44.519 --> 00:32:46.660
doing a lot of research into that and i might
00:32:46.660 --> 00:32:48.819
be potentially getting involved in some research
00:32:48.819 --> 00:32:50.859
in bioacoustics in the future which would be
00:32:50.859 --> 00:32:53.420
very exciting but yeah i think that's a really
00:32:53.420 --> 00:32:56.470
interesting emerging field being able to track
00:32:56.470 --> 00:33:00.710
the location and movements of whales for conservation
00:33:00.710 --> 00:33:04.650
purposes and for notifying ships of when there's
00:33:04.650 --> 00:33:06.690
whales in the area and that kind of thing. Cool.
00:33:07.009 --> 00:33:09.490
Yeah, we had someone on to talk about birds at
00:33:09.490 --> 00:33:12.069
one point and being able to identify species,
00:33:12.470 --> 00:33:15.450
although not yet individuals, from recordings
00:33:15.450 --> 00:33:19.029
just out in the trees. So that's cool. Is that
00:33:19.029 --> 00:33:22.190
like a citizen science thing? Exactly, yeah.
00:33:22.190 --> 00:33:23.730
There's a couple of different citizen science.
00:33:24.509 --> 00:33:27.609
organizations that work on that and again it's
00:33:27.609 --> 00:33:30.009
using neural networks like we have in this study
00:33:30.009 --> 00:33:33.450
to identify what whale vocalizations look like
00:33:33.450 --> 00:33:35.549
and it requires that kind of building of the
00:33:35.549 --> 00:33:39.430
training data for the predictive models. Is there
00:33:39.430 --> 00:33:42.250
any group or anyone that might be listening in
00:33:42.250 --> 00:33:43.910
the Vancouver area that they should get in contact
00:33:43.910 --> 00:33:46.369
with if they want to join this citizen science
00:33:46.369 --> 00:33:49.829
outreach? Yeah, there's a really cool organization
00:33:49.829 --> 00:33:53.440
based in in Washington called Orca Sound. And
00:33:53.440 --> 00:33:55.880
they're the main citizen science people working
00:33:55.880 --> 00:33:58.420
on predictive models. And then there's a couple
00:33:58.420 --> 00:34:00.519
more. The rest are kind of tied to universities.
00:34:00.779 --> 00:34:04.559
So there's a research team at SFU called HALO.
00:34:04.900 --> 00:34:10.119
And then there's also some guys at UBC and UVic
00:34:10.119 --> 00:34:12.320
as well, all kind of working. They all kind of
00:34:12.320 --> 00:34:14.139
work together. It's all a very collaborative
00:34:14.139 --> 00:34:17.039
approach. So I think Orca HALO. The way science
00:34:17.039 --> 00:34:19.900
should be. Yes, exactly. I think Orca Hello is
00:34:19.900 --> 00:34:23.159
the main citizen science kind of approach. All
00:34:23.159 --> 00:34:26.260
right. Now for the traditional last infamous
00:34:26.260 --> 00:34:29.079
question. What is your favorite science joke?
00:34:29.539 --> 00:34:31.820
Yeah, I've got one for your longtime listeners
00:34:31.820 --> 00:34:37.260
that ties into the paper I mentioned earlier.
00:34:37.440 --> 00:34:42.300
So my joke is a Spinosaurus walks into a bar.
00:34:42.840 --> 00:34:47.079
The bartender asks him, why the long sail? Spinosaurus
00:34:47.079 --> 00:34:52.429
says, you know, I have absolutely no idea. The
00:34:52.429 --> 00:34:56.630
joke is that we have no idea what Spinosaurus...
00:34:56.630 --> 00:34:59.730
What it used to say. What it did. I got it now.
00:34:59.789 --> 00:35:02.429
Yeah, yeah. Takes a minute. Oh, boy. Normally,
00:35:02.469 --> 00:35:07.809
I get it right away, or I feign polite indifference.
00:35:08.550 --> 00:35:11.110
This time, I was like, no, no, I'm going to ask
00:35:11.110 --> 00:35:12.969
because I really need to know. Okay, well, thank
00:35:12.969 --> 00:35:15.199
you so much for being on. Yeah, thanks, Jeff.
00:35:15.300 --> 00:35:17.780
It was a pleasure to be on, and I'd love to come
00:35:17.780 --> 00:35:20.059
back sometime. Fantastic, but you'll need a new
00:35:20.059 --> 00:35:22.960
joke. Yes, one that works immediately this time.
00:35:23.179 --> 00:35:25.500
Well, that brings another episode to a close.
00:35:26.019 --> 00:35:28.739
Thanks for joining so far. Remember, send some
00:35:28.739 --> 00:35:31.440
feedback via the email in the show notes. Reach
00:35:31.440 --> 00:35:33.699
out on social media. Make sure you've subscribed
00:35:33.699 --> 00:35:35.820
to the show so you don't miss an episode. And
00:35:35.820 --> 00:35:39.000
give the gift of science. Tell a friend. And
00:35:39.000 --> 00:35:46.739
I'll see you in two weeks. so chaotic so misbehaved
00:35:46.739 --> 00:35:53.780
echoes of melodies remind us it's always colors
00:35:53.780 --> 00:36:01.559
weave stories painting the sky swaying to rhythms
00:36:01.559 --> 00:36:21.059
as the galaxies fly by Deep in our bones Through
00:36:21.059 --> 00:36:28.260
this melodic maze Our minds explore Change