From discovery to decision: the gap Professor Jef Caers is walking toward
- 6 jul
- 9 min de lectura
Actualizado: 10 jul
CAIDTech Research Series | Lima, Peru | World Mining Congress | July 2026
At WMC 2026 in Lima, the Stanford researcher whose algorithm guided one of the largest copper discoveries of the past decade left one question deliberately open. This article takes it from there.
THE PROBLEM THAT DOES NOT END AT THE DRILL HOLE
There is a point in mineral exploration when uncertainty collapses: the drill goes in, assays return, and the hypothesis is confirmed or rejected.
The intelligent agent Professor Caers built with KoBold Metals was designed to reach that point faster and with fewer holes. Not by drilling a fixed grid, but by testing geological hypotheses until an informed go-ahead or walk-away decision is possible. In 2024, the Financial Times reported KoBold’s ultra-high-grade copper discovery in Zambia’s Copperbelt as the first world-class deposit aided by AI-guided drilling [1].
At WMC 2026, Professor Caers put it plainly: “KoBold used our algorithm and made this huge discovery of copper.” Then came the more consequential line: “Now we are starting to look at mine planning together” [2].
That sentence begins with this article because the gap between discovering a deposit and deciding to build a mine is where the industry’s methodology has not kept pace with its ambitions.
Once the deposit is confirmed, uncertainty shifts rather than disappears. Someone must still decide whether to build, when, and at what cost. Under commodity prices, capital costs, operating conditions, geological variability, and regulatory frameworks that may change over decades.
The Intelligent Prospector answers one question well: where to drill next to reduce geological uncertainty. It cannot know the copper price in 2041, predict future royalties, or determine whether an NPV at exploration close will survive the final investment decision.
"Now we're starting to look at mine planning together."
Jef Caers, WMC 2026
RESEARCH NOTE · CAIDTECH Exploration and investment decisions share a mathematical structure but face different uncertainties. Exploration uncertainty is geological (location, volume, and grade) and can be reduced through data. Project valuation uncertainty is economic and institutional (prices, costs, and regulation) and cannot be resolved by drilling; it requires a different analytical framework. |
SEQUENTIAL PLANNING UNDER UNCERTAINTY: WHAT PROFESSOR CAERS BUILT AND WHAT IT IMPLIES
The gap matters because of what Professor Caers built at Stanford Mineral-X: a sequential planning framework in which each decision changes the next, allowing the system to choose the best action based on the information available at that moment [3].
In exploration, the system plans a sequence rather than a fixed drill program. Each hole is guided by the previous result, with the aim of reducing geological uncertainty efficiently and stopping once the remaining uncertainty no longer justifies more data.
That differs from conventional exploration, which drills a fixed grid, builds a block model, and then applies a cutoff grade. Professor Caers’ system integrates data collection and decision-making, placing each hole where it most reduces decision uncertainty per dollar spent.
The results are documented: Intelligent Prospector, published in 2023 and extended in 2024 for epistemic model uncertainty, helped characterize KoBold’s ultra-high-grade copper discovery in Zambia, where direct surface observation is limited [1][4].
Formally, this is optimal sequential decision-making under epistemic uncertainty: uncertainty caused by incomplete knowledge and irreducible through better data.
Investment decisions add aleatory uncertainty: irreducible randomness from commodity markets, geopolitics, and long-term costs. No drill campaign can reveal the copper price in 2041; the challenge is deciding under uncertainty that cannot be eliminated.
That is why the exploration framework should be extended to investment decisions. The question shifts from where to drill next to when, at what price, and under what conditions a project moves from option to obligation; the mathematics are related, but the inputs and horizons differ.
The investment decision problem involves both epistemic and aleatory uncertainty. No drill campaign eliminates the second kind. What changes is not the uncertainty itself, but the framework for making decisions in its presence.
RESEARCH NOTE · CAIDTECH Exploration and project valuation are both Bayesian, sequential decision problems. In exploration, each drill hole updates the geological model and the agent controls when and where to gather data. In valuation, markets update the economic model continuously; the decision is whether to act or wait. Real options analysis, which CAIDTech applies to mine valuation, captures the value of waiting, acting under better conditions, and avoiding premature commitment. It is sequential planning under uncertainty applied to investment decisions rather than drilling plans. |
WHAT PROFESSOR CAERS SAID IN LIMA
Professor Caers’ headline WMC 2026 claim was that AI could cut exploration drilling fivefold, saving time and capital while improving go-or-no-go decisions [5][6]. The gain lies less in algorithmic novelty than in a different logic: drill to test hypotheses, then stop when more data would not change the decision.
His more important point came during the June 25 Talent Focus panel: AI-era engineers must learn not only how to solve problems, but why those problems should be solved [7]. That is methodological, not pedagogical. And it applies directly to investment decisions.
Mining has built stronger tools (block models, simulations, optimization engines) but has been slower to ask whether they solve the decision that matters. Professor Caers’ WMC short course made the same point: use AI to improve decisions through uncertainty quantification and robust data science, not deterministic thinking [8].
In mine project evaluation, deterministic thinking mirrors fixed-grid drilling: both reduce probabilistic systems to a single expected value and can make decisions look rigorous without being decision-relevant. Capital-project research shows the cost of that mistake: downside risk, overruns, and delays, are regularities, not exceptions [9].
Deterministic thinking in mine project evaluation is the same failure mode as drilling on a fixed grid regardless of what the data says.
WHERE THE FRAMEWORK HAS BEEN TESTED
The best-known case linked to Professor Caers’ work at WMC 2026 is KoBold Metals’ use of the Intelligent Prospector algorithm at the Mingomba deposit in Zambia’s Copperbelt. Described as ultra-high grade, at 5% copper or more, it has been reported by Bloomberg, the Financial Times, and Vanity Fair as the first world-class discovery meaningfully aided by AI-guided drilling [1][2].
Less noted is that Mineral-X has applied the same framework beyond mineral exploration. In a 2024 collaboration with OMV, it developed AI-driven tools for geothermal systems and carbon capture and storage, reducing uncertainty while lowering field design, monitoring, and backup-infrastructure costs [12].
These applications show that Intelligent Prospector is not limited to mining. It addresses any geological or physical system that can be framed as sequential planning under uncertainty, where each action changes the next information set and premature commitment is costly. Mineral exploration, geothermal development, CO2 storage, and mine planning all fit that structure.
A 2025 Mineral-X paper extended the Partially Observable Markov Decision Processes (POMDP) approach to critical mineral supply chains, applying it to U.S. lithium resources and showing that sequential decision-making outperforms static planning under comparable conditions [13].
Together, these cases demonstrate the portability of the method: the same POMDP-based structure has been tested in copper exploration in Zambia, geothermal development in Austria, carbon storage, and lithium supply chain management in the United States. The geology, commodities, and timelines differ; the decision framework remains constant.
The remaining gap is the final investment decision. As Professor Caers noted at WMC 2026, KoBold is now beginning to examine mine planning, where the uncertainty shifts from where the ore is to what it will be worth under market conditions no algorithm can predict and no drill campaign can resolve.
RESEARCH NOTE · CAIDTECH The OMV collaboration shows the framework’s reach beyond exploration: geothermal well placement, development timing, CO2 injector design, monitoring, and leakage-risk management. The point is broader: the same logic applies wherever long horizons, irreversible commitments, high failure costs, and unresolved uncertainty make timing as important as action. |
THE FRAMEWORK THAT CONNECTS THE TWO ENDS
Professor Caers has spent the past decade applying sequential planning under uncertainty to mineral exploration. His framework is peer-reviewed, tested across domains, and now being extended into mine planning through KoBold Metals, as he noted at WMC 2026.
During that same period, CAIDTech has applied Quantitative Risk/Compliance Analysis (QRCA) and real options frameworks to mine project evaluation through iEPPS, DAQRAS, Expanded NPV, and probabilistic portfolio selection. The common principle is clear: expected values are not enough; mine decisions should be treated as options under changing conditions, avoiding the “flaw of averages” in project evaluation.
The connection is structural. Real Options Analysis and Sequential Planning under Uncertainty both come from stochastic optimization and ask the same question: how to decide now when better information will arrive later and waiting has a cost.
In exploration, the answer is Intelligent Prospector: acquire data until another drill hole no longer justifies its cost. In mine valuation, it is real options analysis: value the flexibility to defer, expand, or abandon as conditions change.
What Professor Caers is beginning to explore with KoBold in mine planning is the same problem CAIDTech addresses in project evaluation: a rigorous, probabilistic, decision-focused framework that treats uncertainty as the central variable rather than noise to be smoothed into a base case.
Professor Caers’ factor-of-five drilling reduction proves the method. The same logic that rejects fixed-grid drilling also challenges single-point NPV: both are poor ways to manage uncertainty. The mathematics is the same so is the institutional resistance.
RESEARCH NOTE · CAIDTECH The convergence of these methods suggests that sequential planning under uncertainty can extend across the mining value chain, from exploration to final investment decision. Each stage has different data, uncertainty, and timing, but the logic is the same: reduce decision-relevant uncertainty efficiently and commit only when it is low enough for the decision at hand. For critical-minerals projects, this means linking exploration confidence with investment uncertainty to show when a project moves from viable option to executable obligation. A single-base-case DCF cannot do that; it requires the same probabilistic discipline both Professor Caers brought to drill planning and CAIDTech brought to mine Project evaluation. |
ANALYTICAL CONCLUSION · CAIDTECH The most significant contribution of Professor Caers to mining at WMC 2026 was not the factor-of-five reduction in drilling. That result is real and well-documented across multiple application domains, from copper in Zambia to geothermal energy in Austria to lithium supply chains in the United States. The more significant contribution was the implicit argument that the methodology underlying those results, the rejection of single-point determinism, the embrace of sequential planning under uncertainty, the discipline of asking precisely which problem a given tool is solving, needs to be extended to the rest of the mining value chain.
The investment decision is where that extension matters most, because it is where the most capital is committed under the least rigorous analytical framework. A deposit confirmed by an intelligent agent can still be the basis for a final investment decision made with a single deterministic NPV calculated at a single base case price. That is not a failure of the agent. It is a failure of the decision framework that follows it.
CAIDTech is working on the other end of that problem. The tools exist. The mathematical frameworks are mature. What the industry needs is the institutional commitment to apply them with the same rigor that Professor Caers applied to the exploration phase, and the willingness to ask, as he put it in Lima, not just whether we can solve the problem, but why we need to solve it, and whether the method we are using is actually solving it. |
REFERENCES
[1] Mern, J., Corso, A., Burch, D., House, K.Z. & Caers, J. (2024). Intelligent prospector v2.0: exploration drill planning under epistemic model uncertainty. Geoscientific Model Development, arXiv preprint arXiv:2410.10610. https://arxiv.org/pdf/2410.10610
[2] Caers, J. (2025). Stanford's Jef Caers Drives AI Innovation at Mineral-X to Accelerate Critical Minerals Discoveries. Investor News, interview at PDAC 2025. https://investornews.com/critical-minerals-rare-earths/stanfords-jef-Caers-drives-ai-innovation-at-mineral-x-to-accelerate-critical-minerals-discoveries/
[3] Stanford Human-Centered AI. (2022). Building Intelligent Agents to Reach Net Zero 2050. https://hai.stanford.edu/news/building-intelligent-agents-reach-net-zero-2050/
[4] Mern, J. & Caers, J. (2023). The Intelligent Prospector v1.0: geoscientific model development and prediction by sequential data acquisition planning with application to mineral exploration. Geoscientific Model Development, 16(1), pp.289-313.
[5] WMC 2026 / IIMP. (2026, February 24). Artificial Intelligence Could Reduce Drilling by a Factor of Five, Says Jef Caers, Stanford University Expert. World Mining Congress 2026 Official Webinar Series. https://wmc2026.org/artificial-intelligence-could-reduce-drilling-by-a-factor-of-five-says-jef-Caers-stanford-university-expert/
[6] Canadian Mining Journal. (2026, February 25). AI could cut mineral exploration drilling by fivefold, says Stanford's Jef Caers.
[7] WMC 2026 / IIMP. (2026, June 25). AI is changing the rules of professional education, Stanford professor says. World Mining Congress 2026.
[8] WMC 2026 / IIMP (2026). Artificial Intelligence in Mineral Exploration: Separating Hype from Reality. Short Course for C-Suite Decision Makers. Official Programme. https://wmc2026.org/short-courses/
[9] Zangeneh, P. (2026, June 25). Project Risk Management and De-Risking Strategies. Short Course, World Mining Congress 2026, Lima. [Reference to documented research on deterministic frameworks and capital project performance in mining.]
[10] Cobalt Institute. (2026, February). AI: The Road to an Equitable and Sustainable Cobalt Supply. Report authored by Jef Caers, Stanford University.
[11] Stanford Mineral-X. Scientific Research.
[12] OMV. (2024). From geothermal to CCS: how AI is powering our journey to net-zero. OMV Magazine, Collaboration with Stanford University.
[13] Arief, M., Alonso, Y., Oshiro, C.J., Xu, W., Corso, A., Yin, D.Z., Caers, J.K. & Kochenderfer, M.J. (2025). Managing Geological Uncertainty in Critical Mineral Supply Chains: A POMDP Approach with Application to U.S. Lithium Resources. arXiv preprint arXiv:2502.05690. https://arxiv.org/abs/2502.05690
Note: All direct quotations attributed to Jef Caers in this article come from verified public sources: the official WMC 2026 webinar series (February 2026), official WMC 2026 congress coverage (June 2026), the InvestorNews interview at PDAC 2025, and the Intelligent Prospector v1.0 and v2.0 papers. CAIDTech was present at the 27th World Mining Congress, Lima Convention Center, June 2026.
CAIDTech - Centro Avanzado de Investigación y Desarrollo de Tecnología Minera
www.caidtechnology.com · luis.martinez@tmsac.net · Lima, Peru
