- Cerebral Valley
- Posts
- Terra AI is Shaping Our View of Earth’s Subsurface 🌏
Terra AI is Shaping Our View of Earth’s Subsurface 🌏
Plus: Co-Founders John and Anthony on why mineral exploration requires an AI-first approach...
CV Deep Dive
Today, we’re talking with John Mern and Anthony Corso, Co-Founders of Terra AI.
Terra AI is building an AI system to solve one of the most technically complex and under-discussed challenges in the energy transition: how to accurately understand and develop the Earth’s subsurface. As the demand for copper, nickel, lithium, cobalt, and other critical minerals accelerates, the core constraints aren’t refining capacity or a lack of deposits– it’s exploration and development. As John and Anthony detail below, exploration timelines routinely stretch over a decade, driven by noisy signals, incomplete geological data, and manual, high-stakes decision-making. The existing industry tech-stack and approach can no longer scale to the urgency or complexity of this moment.
Formed out of Stanford’s research ecosystem, Terra AI is taking an AI-centric approach to tackling this problem. Today, Terra’s platform integrates generative modeling, surrogate simulation, and decision-making agents to both model the Earth’s subsurface probabilistically and suggest what to do next. Terra maps what could be underground, and then quantifies uncertainty, simulates outcomes, and recommends next steps like drill placement or survey strategy. Their platform is already being piloted by large mining companies such as Rio Tinto, large energy firms, and independent explorers—all looking to improve critical mineral project outcomes and minimize risk.
In this conversation, John and Anthony explain why mineral exploration requires an AI-first approach, how AI architectures are powering their core exploration technologies, and what it takes to build trust in a sector that’s often slow to adopt the fruits of innovation.
Let’s dive in ⚡️
Read time: 8 mins
Our Chat with John and Anthony 💬
John and Anthony, welcome to Cerebral Valley! John, let’s start with you. First off, give us a bit about your background, and what led you to co-found Terra AI?
Hey there! My background is originally in aerospace engineering—about as far from the subsurface as you can get. I spent a couple of years building drones and satellites at Boeing in a group called Phantom Works, where the focus was a lot of prototyping and autonomy. I was building autonomous vehicles from bits and pieces already available, but I really wanted to get ahead of the curve on autonomy and decision-making, so I went to Stanford in 2016 and joined the Intelligent Systems Lab to focus on that research.
I spent a few years working on how to take a perception system—with a few different sensors—and build it into an agent onboard a vehicle that can make decisions the way a human would - like in a self driving car. Eventually, I started applying those methods to a broader set of problems, including energy and Earth resources. The idea was to take that kind of system and apply it to things like designing wind farm layouts for higher output, planning carbon capture well placements to increase revenue and lower risk, or optimizing how we gather data to find critical metal deposits faster.
Even in those early experiments, it was clear we could make a meaningful impact by introducing methods that were familiar to us but new to the space. We found that by applying decision-making algorithms to things like wind farms, we could increase electricity output by 20%. When you think about how much new wind energy is going to be developed in the next decade, that’s a massive deal.
After that, I joined a startup called KoBold Metals, where I got to put a lot of this into practice. Instead of just writing papers, I was actually running drilling programs and directly influencing projects that are going to supply the next generation of critical metals. That was exciting, and by then I was fully hooked on this space. That brings us up to Terra.
Anthony - tell us a bit about your background and what led you to join forces with John.
I met John in grad school—we were both at Stanford, doing PhDs in the Stanford Intelligent Systems Lab. My work focused on robotics and autonomous driving, especially on safety and risk management. I was always interested in how to test and validate AI systems in safety-critical settings before deploying them in the real world. I’ve always thought AI is incredibly powerful, but my two main criticisms are: 1) people often apply it to the wrong problems, and 2) they often deploy it prematurely, without rigorous validation.
When I started looking into problems like discovering critical minerals or developing renewable energy and carbon sequestration projects, I saw huge potential for positive impact—and really hard technical problems. And it was clear no one from the AI world was showing up to solve them. That really excited me, both intellectually and morally.
I stayed at Stanford as a postdoc working on these problems and also served as the Executive Director of the Stanford Center for AI Safety. A big part of what John and I bonded over was the challenge of getting research out of the lab and into the field—taking the things we knew had enormous potential and figuring out how to actually deploy them in the wild. A lot of that motivation came from how difficult it was to make that transition with the limited resources you get in academia. Terra AI was our answer to that.
How would you both describe Terra AI to those who may not be as familiar with what you’re building?
At a 10,000-foot view, Terra is an AI platform designed to help us better understand and develop the subsurface. When I say subsurface, I mean critical minerals, non-hydrocarbon resources, geothermal energy—things like that. The reason mining and these industries are so hard is because everything is underground, gathering new data is really high-cost and there is a huge amount of uncertainty based on what data you do collect. These are often multi-billion dollar projects, so uncertainty is a killer.
Our mission is to solve for the bottlenecks that are preventing us from developing the necessary volumes of copper, nickel, cobalt, geothermal reservoirs — all of which are critical to the net zero transition and beyond. There are plenty of these resources, but they are literally stuck in the ground because we don’t understand them well enough, and we’re struggling to access them efficiently using conventional methods. We believe the way to solve that is by applying the right AI techniques to get a better picture of the subsurface, faster.
Tell us a bit more about what compelled you both to tackle this problem head-on, given your backgrounds. What has the arc of Terra looked like from inception to today?
The thing that’s become very apparent to us is that there is no alternative to solving this problem. When you look at how the critical minerals industry has evolved over the last 20 years, you see massive investment in extraction, processing, and automating fleets—essentially everything at the far end of the chain. We've done so much to improve efficiency in getting material out of the ground, and those gains have been incredible.
Despite all that, the cost of taking a new resource—say, a ton of copper—from deposit discovery to market is now 16 times higher than it was in 1996, adjusted for inflation. So even though extraction and processing have become far more efficient, the overall cost has skyrocketed. And it’s all in discovery and development. Everything before we start pulling rock out of the ground is what’s driving that increase, and the trend isn’t slowing down.
We’ve also invested heavily as a society in new data collection methods—gathering more and more data—but it hasn’t fixed the problem. It’s millions of data points, being analyzed in the head of a human being, who comes up with one concept. These resource models are inaccurate 80% of the time.What motivates me is the belief that the only way we’re going to solve this is by actually tacking the problem head on. We need to move past increments to figure out a fundamentally better way to understand the subsurface.
What’s exciting is that we’re at a point in the AI cycle where, for the first time, certain technologies have come online that allow us to process this enormous volume of complex data—data that’s just too hard for human geologists or geoscientists to interpret effectively. But new AI methods can build a coherent picture of what the subsurface might hold. The convergence of these two factors—an urgent need and new technical capability—makes this a very special moment.
Tell us about your key customers today. Which sectors of the economy would you say are finding the most value in the technology you're building at Terra?
The breadth of customers that have engaged with us has been pretty staggering. We've had pilots with an energy company that's working on an offshore carbon storage reservoir. They're trying to pivot away from hydrocarbons and realizing the traditional oil and gas development playbook doesn't apply here. There are risks unique to this space that require entirely new methods to solve.
We’ve also worked with big mining companies that are facing similar challenges. They’re faced with either working at lower grades at the surface—essentially trying to do more with what they already have– or looking for new, deeper and more complex geological systems. One of the key challenges here is screening– do we double down on exploring this target or walk away before spending too much? And then, of course, developing the exploration targets we do like much more efficiently.
We also work with independent explorers, which are especially interesting. These are companies that don't actually produce metal; their job is to find deposits, measure them, and sell them to developers. They do the vast majority of exploration—about $9 billion out of the $12 billion spent annually—but they don’t have in-house R&D teams. They don’t have the ability to innovate on their own.
So for us to come in and say, “We can take this deposit measurement process—which usually takes about 12 years—and help you get there years sooner, with less risk, less money spent, and better decision-making up and down the line,” that’s a game-changer. That could mean hundreds of millions of dollars in asset value for a company that might have a $30 million market cap to begin with and is diluting itself on the TSX every year. The value of giving explorers and miners access to a state-of-the-art platform—something they’d otherwise have no way of building themselves—is enormous.
Mining could be described as a legacy industry. How challenging or otherwise has it been to impress the vision of Terra’s platform on some of the largest, most established companies in the world?
I'll start by saying the attitude in this space toward AI has shifted massively over the last three to four years. When I first got involved, AI was still seen as vaporware—a passing fad. Now, there’s widespread recognition that it’s inevitable. Some people are genuinely excited to engage, while others are more motivated by fear—this sense that if they don’t adopt it, someone else will and they’ll be left behind. So overall, there’s still a lot of FOMO in the industry, but now it’s paired with real urgency and interest, which definitely wasn’t there before.
The conversations we’re having today are no longer about convincing people that AI is useful. That ship has sailed. Now, it’s more about educating them on how to implement it the right way—and helping them avoid paths that probably won’t work. There’s a lot of noise in the space, with new approaches popping up that we don’t think are viable. So our job is about education, demonstration, and then showing them why we’re the right ones to help solve these problems. We’re way past the point of just selling the concept.
One of the reasons we’ve been met with a lot of excitement is that we don’t approach these engagements like a traditional vendor. We’re not just trying to sell an off-the-shelf product. We show up as collaborative partners. We bring geological expertise, geophysics expertise, and we take a holistic approach to each project. That approach has been really refreshing to many of our customers because we’re not coming in saying we know everything—we’re coming in with tools that help them do their jobs better, while learning from each other along the way.
Having some strong proof points has also helped. There are a lot of new entrants in this space—people fresh out of school, or folks with a promising paper, trying to sell a targeting system or sensor they claim has 99% accuracy. This industry has seen that play out before, and it rarely works in the field. The fact that we’ve already delivered in academic settings and with real customers—like our work in copper mining and carbon capture—gives us a level of credibility that really matters.
That credibility was a big part of what helped us land early partnerships with major companies like Rio Tinto, who’s now not just a customer but also an investor. Having that track record of actually delivering results, not just selling ideas, has been critical in setting us apart with them and other majors.
Walk us through the broad functionality of Terra’s platform. Which use-cases should new customers expect to find value in first, and how easy is it for them to incorporate Terra’s technologies into their existing workflows?
We really do see ourselves as a partner. We don’t think the right way to deliver value in this industry is by throwing a black box at people and saying, “Give us your data and we’ll get back to you with an answer.” That just doesn’t work. We try to be super collaborative, wherever our partners are in terms of understanding, data prep, and everything else. We'll take their data, feed it into our pipeline, and iterate together on what we’re producing.
There are two big components to our system: the generative modeling layer and the decision-making agent. The generative modeling layer typically starts with a customer giving us some pre-collected site data. They may be thinking about collecting more but aren’t sure what makes sense yet. Our team doesn’t just look at the data—we also talk to their geo teams to get all the contextual site knowledge that isn’t easily encoded in structured data but is still critical to exploration. We work together to figure out how to represent that information in a way that our pipeline can use.
From there, we generate a set of outputs—sometimes tens of thousands of 3D subsurface models. We can aggregate those and produce summary statistics, or we can zoom in and look at individual models. What’s key is that we sit down with their geologists and show them results in a format they’re familiar with. We ask, “Does this look geologically reasonable to you?” We’re not tossing over abstract prospectivity maps or dense magnetics plots—we’re speaking their language and building trust in the process.
Then we move to the decision-making layer. Once we’re all aligned on the modeling side, the next step is understanding what decisions they’re trying to make. Are they deciding where to drill next? Whether to run more surveys? Or whether to abandon the project entirely because it might not be economically viable? Our reasoning agent evaluates all these options, quantifies the expected value of each, and recommends the best path forward.
To simplify it: the generative modeling piece takes in site data—things like drill core samples or geophysical readings—and produces 3D models of the subsurface. These models reflect both what we know and what we don’t know, capturing uncertainty in a meaningful way. Then our decision-making layer uses those models to simulate different paths, reason through the potential outcomes, and deliver an actionable recommendation.
How are you measuring the impact and results that Terra’s platform is creating for your key customers today? What are some of the areas - speed, cost, quality - that you’re most-heavily focussed on metrics-wise?
One of the really nice things about the decision-making methodologies we use is that we can align our optimization objectives directly with the specific business goals of our partners at any given time. So in the case of a reservoir development project, their top priority was injecting a certain amount of CO2 while maintaining a very high level of safety. They wanted to be highly confident that there wouldn’t be any leakage, while still achieving their injection targets.
For that project, we tracked exactly that—injectability and leak risk. We generated a range of solutions and compared them directly to human-designed ones. What’s powerful is that we weren’t just looking at point estimates like “this design injected X amount of CO2.” We provided probabilistic outputs with full distributions. For example, we could say with 95% confidence how much CO2 would be injected, or with 99% confidence that there would be no leakage. It’s a lot easier to speak directly in the customer’s terms with that kind of granularity.
In general, on the reservoir side, the metrics we care about are improvements in operational performance, cost savings in monitoring, and above all, safety. We’ve seen significant improvements over human designs—double-digit percentage gains in both injected volume and risk reduction.
On the mineral side, it depends a bit more on the stage of the project. One recent project involved a mine site that had been explored for decades and was nearing end-of-life. The team was looking to see if there were new ore bodies still hidden there. We ingested all their historical data and prior analyses and used our tools to search for new targets. We ended up identifying several million tons of potential new resources. If realized, that extends the life of the mine by years and generates a lot of copper that is needed to build renewable energy.
The key metrics there are megatons of new resources discovered, and time savings toward the next decision gate. At large miners, there’s often a rigorous gating process—drill or drop, delineate or abandon, and so on. We’re trying to cut time out of that decision cycle. Today, post-discovery, it takes about 12 years on average to fully measure a resource. Our aim is to shave about 40% off that timeline—whether it’s faster to drill/drop, delineate, or ultimately bring a project to market. That time is worth a huge amount of money.
There are a number of talented teams pushing the frontier of material extraction using AI. What would you say most sets Terra apart from a platform perspective?
The two things that really set us apart are first, the people we have—which directly translates into what we're building—and second, the way we do business, which is fundamentally different from most others in the space.
On the people side, we've made a point to hire deep experts so we can build really foundational work that's unlike anything you can pull off the shelf. Our generative platform is completely novel in terms of what it’s capable of. Combining all different types of data at once isn’t something you can do with an off-the-shelf system, but it's exactly what the space needs. We've been very deliberate in hiring highly specialized researchers so we can build something truly aligned with what the industry demands.
On the business side, we're not a vertical explorer and miner, so we’re able to impact a broad range of programs across the sector. We work with a wide range of explorers, mining companies, and energy companies, and can potentially influence every major project that comes through Terra. We’re not locked into a single node or investment. We’re focused on identifying all of the key value drivers our technology can improve and targeting them, not any of the less differentiated verticals like permitting land and raising development capital.
Lastly, we’re also not a traditional SaaS company. If you want to run a pure SaaS model in this space, you need to sell a simpler product—something easy to use but probably less performant than what the state of the art can actually do. We’re more high-touch than most, but the benefit is we’re always delivering the best that’s technically possible at any given time. We also take the time to deeply understand our customers and build trust, which is critical to getting this adopted in the long run.
Walk us through Terra’s platform and use of AI under the hood. What are some of the key AI architectures, systems and models powering Terra today?
Let’s start with the generative modeling side, since that’s probably the most familiar to your audience. One of the big innovations in recent years has been the rise of deep generative models—specifically diffusion models—that can generate high-dimensional data resembling the examples they’re trained on. Most people know these from image generation. What’s especially powerful is their ability to generate content conditioned on some form of context. In image models, that’s usually a text prompt—“an astronaut riding a unicorn,” for example.
In our case, instead of using a text prompt, we condition the model on site data—things like drill cores, geophysics, or geology logs. These data signals carry a ton of rich, contextual information. While they’re not formatted like text, we can still feed them in as input to the model. And instead of generating 2D images, the model generates full 3D geological models of the subsurface. It’s the same core tech as the image and video diffusion world, adapted to the Earth.
Another cool piece we use is surrogate modeling. In the geosciences, you often need to simulate really complex physical processes—things like injecting CO2 underground, geothermal energy production, or running synthetic magnetics surveys. Traditionally, these simulations take hours or days to compute. But with surrogate modeling, which is a specific application of supervised learning, you can train a model on thousands of those simulation outputs. The result is a “surrogate” that runs 100,000x faster but retains most of the accuracy. That’s been a big breakthrough.
Then there’s the decision-making piece, which builds on reinforcement learning. We’re pulling inspiration from AlphaGo and AlphaZero—algorithms that combine Monte Carlo Tree Search with deep RL to make strategic decisions. We’ve adapted those ideas to geoscience. So the model’s not just making predictions; it’s reasoning about actions and choosing the best move on the “chessboard of Earth.”
One last thing to add here is the origin story. Terra came out of a pretty special collaboration between the Stanford Intelligent Systems Lab, which focused on decision-making and reinforcement learning, and the Stanford School of Sustainability, which was doing cutting-edge work on modeling uncertainty in the subsurface. That cross-pollination is what birthed this approach, and I think Stanford was one of the few places in the world where those pieces could’ve come together like that. It’s a big part of what makes our team and our platform unique.
How do you foresee Terra evolving over the next 12-18 months? Any priorities or product developments that your customers should be most excited about?
The key priorities for us are continuing to deploy the technology across more companies, more projects, and a broader range of commodities. Our vision for Terra is that this becomes the standard way the industry operates. That's really the only way we're going to solve the challenges posed by the upcoming resource shortages—by unblocking progress across every new exploration project.
So far, we've run 4 pilots. Now we want to hit a cadence of 15 to 20 contracts a year, ideally within the next year, and then double that annually. The long-term goal is to be involved in every successful and important exploration project happening globally.
In the short term, our main priority is scaling both the team and the technology to support that vision. We’re focused on going after the 15 to 20 highest-priority resource projects and playing a key role at that pivotal make-or-break stage of exploration and development.
Lastly, tell us a bit about the team at Terra. How would you describe your culture, what do you look for in prospective team members joining Terra?
We have a great team—about 10 people right now—and what's unique about Terra is how multidisciplinary the team is. We've got geologists, geoscientists, and geophysics specialists, alongside core machine learning experts working on deep learning methods for surrogate modeling and generative modeling. We have people focused on reinforcement learning and decision-making systems, plus strong generalist software engineers and ML engineers who help scale our infrastructure and coordinate across domains. We’ve got reservoir engineers—essentially the fluid dynamics experts of the subsurface. Also, we have successful mining investors and operators as well as a few ex-founders.
One of the core aspects of our culture is a growth mindset and open-mindedness. Everyone collaborates across fields, learning from some of the best in those specialties. It makes for a super dynamic environment. We’re hiring now, and we’ll be making a bigger push later this summer. We’re looking for machine learning engineers, software engineers, and people with experience in scaling infrastructure and models into real-world deployments. Our hiring posture is shifting a bit—from initial prototyping and R&D to scaling and productizing.
Beyond those roles, we’re also looking for strong engineering leadership - that includes a VP of Engineering and software leads who know how to build real, production-grade systems. The common thread we’re looking for is a passion for the mission and a willingness to learn across disciplines. We've got PhDs, folks from NASA, Berkeley, and elsewhere with serious domain depth. Everyone here is expected to learn from everyone else - that's the kind of team we're building.
Anything else you'd like our readers to take away about your vision and the company you’re building with Terra?
First, we’re fundraising in June and opening up a Series A round. If anyone’s interested in participating or learning more, we’d love to chat!
Second, we just want to hit the mission one more time. When we first got into Earth resources, we never thought mining would be where we’d land most passionately. But the reason we did is because if you look at any net-zero transition plan, any electrification agenda, or even global industrialization strategies, the biggest constraint in every single one is the shortage of copper, cobalt, lithium, and other critical metals. And there’s no business-as-usual way to get them.
We really believe that accelerating the discovery and development of these resources—opening up this bottleneck—can have a massive impact. Not just for electrification and decarbonization, but for the broader world: helping nations that are electrifying for the first time, or advancing industrial capacity globally. These materials aren’t replaceable, and we can’t scale access to them using outdated methods. We haven’t met a single mining professional that believes existing processes will get us the materials we need in a timeframe that doesn’t imply climate disaster.
If you're serious about climate, about energy security, or about solving foundational industrial problems—there’s nothing more impactful than working in this space…
Conclusion
Stay up to date on the latest with Terra AI, learn more here.
Read our past few Deep Dives below:
If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.