AI Powered Mineral Exploration (Guest JP Paiemont, VRIFY Technology)

Karl Woll: 00:00:00

Hello, and welcome to Rock Talk, where we explore the world of mining through casual conversations with industry experts. I'm Karl Woll, Senior Account Executive at VRIFY Technology. And in this episode, I am joined by Jean-Philippe Paiemont. Director of Artificial Intelligence and Targeting at VRIFY. In this interview, JP talks about his background in mining, including consulting work, using historical geological data alongside machine learning, to identify and visualize drilling targets for companies. JP then gives us an overview of VRIFY's new AI powered mineral exploration program.

As this is a new podcast.. I would love any feedback you have. If you have any suggestions for improvements to the show or guests, that would be great to have on, I will link to a feedback form in the show notes, or you can go to rocktalkpod.com, and there's a form on the website. And with that, let's jump into this week's conversation with JP Paiement

JP, how's it going? Thanks for joining me.

JP: 00:00:56

Yeah. Thank you for for having me.

Karl Woll: 00:00:58

Perfect. Well, I wanted to talk to you today about our new exciting product at VRIFY, our AI-powered drill targeting service, and have you take us through a walkthrough of that. But before we get into all that fun stuff, I wanted to get to know you a little bit about you.. So by, by way of introductions or just to kick things off, JP, do you mind giving us your background a bit about yourself and, and how it is you got into mining?

JP: 00:01:19

Yeah, for sure. I've studied geology at university. So I did my undergrad in Montreal mostly focused around sort of more of a technical degree where it was a lot of field work, rock identification, minerals and all of that. And then I did a masters in metallurgy based out of Laval University in Quebec City. And from there, I actually always worked as a consultant.

I joined a consulting company, almost out of school, which brought me to different projects different environments. Like I've worked in South America North America, Africa, Asia. Australia. So different types of project, different commodities. So I got to see a lot of different types of mineralization, different types of deposit, but also a lot of different data sets and sort of ways to organize all of that data. And what brought me to geology to start with is just being outdoors was just a common denominator of everything I applied to at school. So I was accepted in three different programs, ended up choosing geology just because it seemed like, I guess a future in terms of job security and whatnot, but yeah, just uh, love being outside.

So the adventure side of it as well. So getting to travel to remote places and explore was what drove me in to start with.

Karl Woll: 00:02:39

Do you have a favorite part of the world, like thinking of the outdoors and that adventure aspect of it and your career? When you were traveling the world, do you have a favorite location or spot that you ended up?

JP: 00:02:48

I've got a few actually. I've always been a big mountain person. So whenever I get to see big mountains is something I really like. I guess my highlights would be Mongolia was really interesting in terms of the culture. Scenery is amazing. And then as a geologist, there's almost a hundred percent exposure everywhere.

So you get to see the rocks everywhere, which is just makes things way easier compared to the sort of marshy, boggy Abitibi landscape. Otherwise, Morocco has always been great. Again, culturally, it's been so fun to work in. It's a little bit easier because there's no language barrier. Most people, well, for me, actually, because most people speak French. So being French Canadian helped over there. And then Africa in general. So I've worked in Mali, DRC, Uganda. So it's always been sort of the laid back style of working over there has always been appealing to me, once you set aside your North American sort of principle of being on time for everything, and that's just sort of generalities, but just the less stressful environment is something that I sort of enjoyed as well.

Karl Woll: 00:03:49

And with your early career and you said doing consulting work, you've been with VRIFY for, about a year now or so working on our drill targeting platform. But thinking back to those consulting days I know you have a background with our CEO, Steve De Jong, and the company he was running at the time, Integra Gold. And the Integra Gold Rush Challenge. Could you walk us through what the Integra Gold Rush Challenge was and your participation in that?

JP: 00:04:17

Yeah, for sure. Yeah, in terms of connection, though, I knew Steve before joining sort of participating into the Gold Rush Challenge. Actually, I've I was a geologist from managing a drilling campaign for him and a past company in the Timmins area. So that goes back to even longer than the Gold Rush itself. And then, yeah, Gold Rush Challenge was sort of a contest that he put together where they had a hefty data set that didn't really want to sift through to try to find targets for an area that was outside of their main focus area. So what they did is they made the data set public, you could download it and then basically had to submit a technical targeting submission.

So come up with exploration targets I think it was like a 10 pager document that you had to summarize your work in and yeah, you had to submit that and they chose, I want to say the five best for a stage sort of Shark Tank style final. Yeah, so it came in a time where consulting was a bit low on work just because it was uh, during a downturn in the mining industry.

So obviously we had a little bit more free time at the office. So we decided to undertake that work. We sort of approached it traditionally while sort of integrated interpretation has always been something that I do a lot of. So it's like getting all of the data into a 3D model and then interrogating that model to come up with targets.

So initially that's how we set it up. And right at the end, we decided to add some machine learning to it, just to see if we could sort of, yeah, generate new targets using the data itself without building in the uh, geologist's biases. We ended up. selected for the final, went on a final on stage, explain everything. And uh, we ended up winning the challenge. So it's sort of sec... For me, it sort of was a catalyst to um, learn more about AI and how it can be applied to the geosciences that really got me the exposure to that. And yeah, just the event in general is quite, quite fun to, to, to be a part of.

Karl Woll: 00:06:09

Yeah, I find that challenge like really interesting because obviously in the industry, data is everything and most mining companies keep their data close to the chest and it's proprietary. So taking the completely opposite approach and just saying, here's our data set, tell us where we should be poking holes in the ground for shiny metal is is an interesting approach.

When they were judging the finals, like what was, how were they grading you? What were they judging you on? And, and how was that decision made in terms of determining a winner?

JP: 00:06:35

I on the technical side, like there was a big aspect on the technical side when I think that's choosing the five finalists was fairly technically judged. So basically they had a panel of industry and academia leaders to sort of. decipher I want to say 2000 and that's just me throwing out a number there, but I want to say about 2000 submissions that they received. So they had, yeah, like a heavy technical side to it and that part of the of the judging and then the onstage was five sort of sharks. So it was more, more of a, like your business approach to it, your vision, how well you could pitch in five minutes or seven minutes, I think we had. So you had to be sort of concise and also be able to convey your point across pretty quickly. Yeah, so the Shark Tank panel was mostly CEOs, CFOs of different companies, so more on the business oriented side for the finals.

Karl Woll: 00:07:28

And I've seen a photo of you with a oversized novelty check. so I think it was like a half a million dollar prize that your team won. Any any fun stories with the money?

JP: 00:07:38

Yeah, well, for sure, I think the night of, I carried the check around town in Toronto, tried to bank it at different bars. They wouldn't, they wouldn't change it for cash, obviously. But we had, we had a quite a good time. But yeah, so we just walked around town with a big check. Unfortunately, or fortunately for my company, we did all of the work on their company time. So the money wasn't ours to keep for the full amount. It came at a good time for the company. And could just, again, being in a downturn, getting a half a million dollar revenue off of a very little work was pretty good. Yeah.

Karl Woll: 00:08:09

Very little work. I'm sure Steve will love to hear that part of it.

And that's a good segue to I guess present day. So you've done the consulting work. You have that background going back. You know, I think the Gold Rush challenge was maybe eight years ago. So you mentioned machine learning and leveraging data models to help companies find mineralization on their projects.

Can you take us through your role now with VRIFY what you've been up to the last little while and your day to day right now?

JP: 00:08:34

Yeah, definitely. I was brought into the VRIFY team November 2023. Ever since I've sort of focused my attention on building a way to sift through all of the existing data that clients have or that company have to establish a predictive model on to where to find mineralization next.

So basically we're learning from areas they know. And through their data, the exploration data they already have, trying to decipher what constitutes the signature of the deposit, what's the mineral system, how the mineral system is represented in the data. And from this, we make a prediction on to where we should find the same type of signature on the property.

ch, which happened at PDAC in: 2024

Karl Woll: 00:09:49

Awesome. Well, given VRIFY is a visual presentation tool. Can you give us a quick walkthrough of, you know, exactly what, what we've built out here, what those visualizations look like and how those predictions look in in practice. So once we've received some data, what is the model actually doing?

What are those data sources and what are the outputs of the model?

JP: 00:10:09

Yeah, for sure. Just to start with the way we're intending it and you like we've mentioned consulting a lot in our discussion right now, but we're seeing that the whole approach we're having with predictive modeling here at VRIFY, not as a consulting approach, but mostly as a software solution.

So basically, you would have access to a cloud computing platform to ingest your data and run predictive models yourself and see the results. And that all ties up into the VRIFY presentation tools that that already exists, which has a really strong and compelling visual aspect to it. So yeah, if we look at the screen, we have our toy data set right now.

So that's just a Flin Flon, Manitoba. The reason we chose it. It's just because there's a healthy amount of publicly available data over top of that area. But It's also an historic, producing area. So we can sort of validate our prediction against the reality of the ore body that was actually mined.

So if we're able to sort of predict the location of an ore body and it matches what we already what was mined in the past, then we know that we've done a pretty good job. So, Yeah, so this is the area here. What we do first is compile all of the exploration data that the company has, or that area has into what we call the data stack.

So the data stack is just a 2D grid representation of all of the data that's available for exploration. So most of the time it represents things like geophysical surveys we'll build a distance factor to different interpreted geological objects. We'll have things like soil, geochemistry maps, we'll have the rock geochemistry maps. Any sort of exploration data or exploration vectors that the company has will build into those grids and then ingest it through the platform.

So that data stack can be 2 layers high, and it could be 50 to 70 layers high, depending on how much data exists over the over the top of the property. One thing we really believe in as well is there's a lot of, probably dormant mineralization or deposits that sit in those data stack that people just haven't identified because of the way they've been either treating their data, because of the cognitive bias that they have in their exploration program.

And we feel like just using AI was will be able to unlock a lot a lot of this value. So once we have that compiled, what we'll also do is compile, the drill holes or the surface geochem samples. But basically, we compile data that's able to help us train the model. So train the machine learning algorithm to make the prediction.

So what we use in that case is the drill holes. So for this one, we've sort of created the drill hole database to remove all of the actual exploration holes that hit the massive sulfide. Because what we want to do is basically see if we can find the massive sulfide lens through predictive models. We only keep a small part of the ore body and that's what you're seeing here.

So what we did is kept the drill holes that intercepted the massive sulfide lens there and our approach here with that toy data set. It's just your typical exploration problem where you've hit some mineralization somewhere in your drilling and you're trying to see where it goes from there.

So are you going to drill down the direction? Are you just drilling down that direction? Trying to find it there? Or is it actually repeating itself somewhere else, right? So that's your sort of typical exploration. You have some mineralization, you're just not sure where it goes from there and you're trying to find that out. So instead of trying to blindly test or sort of not blindly test because most, most exploration program aren't blind, they're educated decisions that we make uh, can we find a way to build quantitative metrics into that decision making process, right? Can we have more of a probabilistic approach to exploration modeling? So So we do is from those drill holes, we'll create what we call learning examples and these learning examples are areas that we know are mineralized or not mineralized. So basically, that means that we convert all of the drill hole traces into either a positive example or a negative example.

So if we take one drill hole, for example, here, if we've hit copper or zinc here in that drill hole, that becomes a positive and basically the rest of the drill hole just becomes a negative example. So now the machine knows where in the 2D space where on that map, is actual mineralization or where is no mineralization?

And we've tested for, right? The rest is just unknown. And what we're trying to do is basically learn from the signature of the positive and negative through the data stack and make a prediction for the rest of the unknown areas in the model. So what we'll do is we'll select an area, a cluster of drilling like this one here and we'll train our model and that model looks through the data stack trying to make a prediction for the cells that we know the outcome. Once the model is trained enough, we'll go and apply it on another area, and then that allows us to test whether the prediction we're making makes sense or accurate or not, right? Since we know some of the results there, we can say, well, 89 percent of the time prediction we make for the model is the same as the actual known point for that general area. And then we just shuffle these areas around.

They're just to build some more resilience into the model. Ultimately what we output is a prospectivity score. So what the prospectivity score actually is is, just a probability of finding the mineral system for every cell in our model. So this is sort of a heat map to where you're most likely to find, in that case, it's a VMS system.

So where is, the area you're most likely to find a VMS system on your map? And it's quite easy to understand, right? So you look at that, you know, where are the areas that are sort of more likely to be mineralized compared to sort of trying to understand a geology map with a mag survey and an EM survey and some geochem points and trying to make that image yourself, in your head. Now you have a product that just makes that image for you. Since we work off of drill holes, we can also sort of push that prediction in 3D. So in the end of the day, you're getting a point cloud. Every point in that point cloud has a prospectivity score attached to it. So we can say, well, these are the areas that are most likely being mineralized. And then you can see sort of clusters of high prospectivity score. These become actual expiration targets. So, for example, we have one target here. So what we could do is group all of the cells that have high prospectivity for that target, and out of that we can estimate different things. We can sort of have an estimation of volume. We can have an estimation of the average score. So what's the probability of this target existing? We also put out the variance on that prediction. That means that we know how risky that exploration target is. So instead of just trusting the geologist that says this is where we should drill. Because I know what I'm doing, and most of the time it's accurate.

You should trust these geos. You're just sort of adding another layer on top, which is more of a quantitative metric and approach to mineral system targeting. The other thing we do is for every target in our model, we'll also output a label. So these labels contain a different types of information there.

So on the left is the feature importance map and the feature importance map would tell you which of the data in your data stack, which of the layers in your data stack are most predictive for your prospectivity scores.

So essentially which one holds the most predictive power. So all of your layers are ranked from lowest predictor to highest predictor. Then, everything to the right of that line here means that it's a good, value to predict high prospectivity. So the way you would read that is your most predictive layer in that model here is the gravity gradient on the YZ axis. That's a type of airborne survey that we fly to find mineral systems. You would be looking for high values in that grid. So if you were to just threshold that grid of data like we normally do in geology, you would just threshold basically filter out all of the low values and look at the high values to try to find your mineral deposit. And then you would combine to that the second best predictor is again the gravity gradient, but this time just on the pure Z component. And for that one, you're looking for low values. So a high in a certain grid combined with a low on another grid is actually making that target appear out of the unknown areas. And you can keep going down the list. Like the third one is a high mag area that's also associated to mineralization. So as much as it is a predictive modeling to sort of help you target mineral system outside of your areas of knowledge, it also sort of serves as a data mining exercise, right? So you can understand how all of your inputs influence your prediction.

And now changing these inputs would change your predictive models. So that's really interesting in terms of the technical team, looking at these these types of of approaches. And then you can start making parallels to your mineral system analysis. And you can also sort of really understand if it makes sense, if the model is not hallucinating or just creating things out of thin air.

So for example, here, we're searching for VMSs, which is a massive body inside less dense rocks. So in that case, it would make sense that gravity is the best predictor because you're looking for gravity high most of the time for these sciences. So you can understand how the model makes its decision. Yeah, I guess lastly, the question is, does it actually work right? So here what you're seeing is just the actual massive sulfide lenses that were modeled out of the circle data and you're seeing the prospect to be score over top of it. And what we're seeing is we're getting a pretty good match between, what is the predictive location of those ore bodies and where they actually are.

And keep in mind that none of that information was fed into the system. We're just using it as a post priori sort of test of the results. So basically, once we've made the prediction, we look at the solids and say, well, does it actually make sense. Does the prediction match what we know in terms of the reality?

So that's why we're using that data set. That's just because we know the reality and we can test the predictions against the reality there. So we're quite a happy with those types of results. Yeah. So in a nutshell, that's how it works. So you're using your exploration data, you're using learning examples from your mineralization and try to predict where you should find that type of realization elsewhere in your property.

Karl Woll: 00:19:53

So that's really interesting. So to your point earlier, where you're saying this isn't really a consulting service, we're not just receiving data from the client, creating a report and saying, here, you should drill these areas. This is a living, breathing model. That the client has access to A, run these iterations as many times as they want with different data layers or adding new data, removing data layers, but it gives the team of geologists an actual output.

So it's not a black box. So they can see what these values are, which layers are the most predictive in mineralization for this particular target. And then maybe that matches existing targets, or maybe it's a completely new target that had been previously overlooked. But it gives the team some data points where they can actually go and have a discussion.

And then to your, to your point, have a discussion. Does this make sense? Does this not make sense? Why or why not? Rather than just blindly trusting some report.

JP: 00:20:45

Yeah, exactly.

Karl Woll: 00:20:46

So, just to clarify what we're looking at here with that, ore body in the red at the forefront of the screen, none of the drill traces, from the dataset that actually pierced that ore body, those were removed from the model. And then when we ran the model, it pretty much matched, or predicted where that ore body should exist.

JP: 00:21:06

Yeah, exactly.

And again, that's what you're seeing here on the screen, right? Where we've trained on only intercepts in that area. And then the prediction sort of swept across the rest of the unknown areas and make that prediction and we can see a good match.

Karl Woll: 00:21:20

And because I'm not a geologist obviously, but I've heard you talk about this a few times before. There's a couple things in here that the model predicted in terms of geological features that I think were interesting. One is the fault, and then one is that there's a wave pattern when you look at it from, from head on.

Can you take us through A, why those are important geological features, but B, what's interesting about the prediction or the output from the model?

JP: 00:21:44

Yeah, for sure. What we're seeing here is basically a side view of the prediction model. So in the terms of a VMS system, it's formed at the bottom of the sea floor. So you're looking for that interface or that old sea floor, basically, to find those VMS's. With time the seas being closed by geological processes and just folds everything up.

So you're basically trying to find something that was flat. That's now folded into tight folds. So what we're seeing in terms of the prediction here is that, that horizon that we're chasing, which is a paleo sea floor, we can see in the point cloud that we're generating. So what we're seeing is that we're sort of generating an undulating surface, which is the folded up, seafloor.

And the other thing we're seeing is that it's broken up in some places. So you see here, there's something what we call a fault. Basically, it's broken up that surface into two different areas, and that's just due to again, if you compress rocks, they're going to break and generate faults. When we look at the actual model, this is the geological model that was built over time for that area. If we put on the faults, we'll see that the model actually made a really good, prediction in terms of where these, faults should be.

So if we light up the faults, what we're seeing is basically out of the exploration data, the model was able to sort of understand the geometry of everything and really predict the location of that fault.

So if you look at the 30 predictions here, you see that even though the faults weren't used as part of inputs to the model, the undulating surface of the paleo seafloor was actually broken up by faults, and we can see that the model matches these faults really, really, really well. We're sort of seeing it pretty much, different spots.

So you saw it from the side view. This is sort of the view from the front. And then again, we're seeing the same thing where there is a fault here, cutting the model, and all of the rocks below it aren't good for VMS system, and we don't get any predictions there. So the model, sort of the way the prediction is made and the spatial distribution of the prospectivity score makes what we call geological sense.

So if you were to look at the results and understand the geology, both of them match pretty closely.

Karl Woll: 00:23:49

That's really cool. And then I guess this is a, a model where we reverse engineered, a known mined, ore body, some publicly available data and built it out as a test case. Where are we or where is VRIFY in terms of rolling this out to to other companies to actually make real world predictions and targets?

JP: 00:24:10

So ever since we launched it, the initiative would have four clients at PDAC. We've signed for 14 additional clients, so all of these people are at various stage of the prediction modeling where what we're working with them is compiling the data, getting formatted and quality checked in order to be ingested by the platform. Some other clients are at the stage where they have a predictive model in their current exploration campaign, we have different clients that are either in different phases of drilling right now. Other clients are currently field prospecting. Some of those targets are quite near getting some real world validation and by real world validation what we mean there is just somebody going and explore one of those AI generating targets and finds mineralization. It could be just a trace of it, but it could also be a new deposit. But really, what we're trying to see is sort of a real world example where if somebody went somewhere, they didn't, thought of going first and and we're able to find something that they didn't expect in that area, right? So that's going to be all ultimate proof of concept. So we, we expect that maybe by mid August, we'll start getting some sort of good validation out of our different clients.

Karl Woll: 00:25:19

It's coming up quick. It's really exciting. I can't wait to see some of the results from our early clients and, and the outcomes of the Truth Machine. So that'll be that'll be a lot of, a lot of fun to watch. Thank you very much for your time today, JP. Really appreciate you taking us through this.

It's a really I just think it's really cool. It's really fun. It's just a really high tech, modern way to approach mineral exploration and what I really like about this and maybe just I'm biased because I don't have a strong background in technical background in mining is from the simplification standpoint from an outsider looking at, you know, Okay, we should target this area because we've compiled a bunch of data, run it through a machine that gave us these targets, and we're going to go drill it and see what happens.

And you can have these visuals that really make it clear where we're, where we're targeting and why. I just really like how it simplifies it and keeps it engaging for someone like me to follow along with the story and just see, it just makes me really curious as to waiting for those assays to come out compared to the typical approach, which is looking at a polka dotted map of some geochem and some geophys, which I have no idea what I'm looking at and have no idea where those holes are actually being poked and what to, what to expect coming out of those.

When it's in a format like this, I find for a, a simple brain like mine, it works quite well. So I really uh, I know there's a lot of technical work that goes into collecting all the data and compiling the machine learning, but for me, I just love the simple visual output.

JP: 00:26:44

And that's part of it as well, right? We're trying to attract people to those companies into that market where we need to make a product that's easier to understand to sort of the general public. And the other thing that we should mention is, as you see those models, once we start collecting results, and once the company start either like drilling or field testing, prospecting, they can easily update their model so you can log on to VRIFY in two months and see how the targets evolve from the initial prediction to the validation to another drilling campaign or flying an additional survey.

So you could see sort of how these targets would evolve, how they take shape, and really start following how the company explores more in real time than waiting a full year for like publicly disclosed results or just, yeah, assay results that are being published on uh, press releases.

Karl Woll: 00:27:31

It's awesome. Well, thanks again, JP. Really appreciate your time. I know you've got a couple weeks off with the family here. Speaking of traveling, I think you're off to Morocco. So, hope you enjoy your trip and look forward to seeing this to to continue to evolve into the near future here.

JP: 00:27:43

Yeah, thanks for your time, Karl.

Karl Woll: 00:27:44

JP. Cheers.