礦產勘探中的AI：2025年回顧

PosGeo

AI in Mineral Exploration: 2025 in Review

2025 was a blockbuster year for both AI and critical minerals! Pretty much every other week either we got news of a new discovery or governmental initiative, or an AI company releasing a new model or product. To break through the noise and try to understand the trends of the year, I thought it would be fun to write a “Year in Summary” of AI in geoscience (with an exclusive focus on mining and subsurface inference), following the pattern set by some excellent yearly summaries I’ve read over the last few days: 2025: The year in LLMs by Simon Willison, and Databases in 2025: A Year in Review by Andy Pavlo.

Disclaimer: The contents of this blog post are headlines pulled from company press-releases, news agencies, and aggregators, woven with my personal takes. Opinions displayed here represent my own work, and not any opinions of my employer.

Let’s get into it!

Funding

In the funding world, KoBold kicked off the year by raising a whopping $537 million in a Series C round (post-money valuation of ~$3B). They claim that funding will be put towards global exploration, R&D, and advancing their Mingomba site towards production. VerAI followed the next month with a $24 million Series B round, promising to accelerate their 60+ mineral projects. Meanwhile, GeologicAI closed a $44 million Series B round, aiming to deliver “high-resolution decision engineering” on drill core, integrating cutting-edge core scanners and proprietary AI to deliver “a new standard in high-resolution data and decision-making” on site. By December, they had also acquired Lumo Analytics, makers of a compact Laser-Induced Breakdown Spectroscopy (LIBS) rock-scanner, adding fast light-element and REE analysis to their sensor suite.

These funding raises are impressive, and give me great confidence for the future of the concept of AI in exploration, both for vertical business models as well as service providers! It’s particularly impressive to see that KoBold can continue to bring in high-value investors like T. Rowe Price, demonstrating both the value of AI and that exploration is open to this class of investors. KoBold claims that AI let them find something that traditional methods missed, but I’m interested to see what will happen next: can they use AI to accelerate development? GeologicAI’s tech stack sounds super cool! Fast access to data should be able to accelerate decision-making, and I’m eager to see them put it into practice.

AI Products

Big AI companies dropped a string of blockbuster products in 2025: reasoning LLM models like GPT5; professional-grade generative image models like Nano Banana; and revolutionary agentic coding frameworks like Claude Code. On the other hand, the AI tech stacks being built for mineral discovery feel like a very different “flavor” of AI. Companies like KoBold, VerAI, and Vrify are mainly solving inverse problems: inferring the unknown subsurface given sparse signals. In press releases they largely reference feature extraction/engineering, pattern recognition, and other pragmatic machine-learning approaches, rather than LLMs and image models. (That being said, the sensor-first, data-fusion pipelines of GeologicAI stand apart from the crowd here.)

That’s not to say that LLMs and generative image models are irrelevant to mining. In December, VRIFY released their 2025 Mineral Exploration Tech Report, claiming that 56% of people working in exploration are using AI tools in their day-to-day work (based on a survey of 135 industry professionals). But it’s one thing to have mineral exploration professionals, say, using ChatGPT to offload tedious tasks like summarizing vendor reports. It’s quite another thing to use or create novel, bespoke AI tools and products. I suspect that many C-suite leaders at these companies are being required to have “AI initiatives” that are (hopefully) more in line with the latter. Exactly how they accomplish this is not immediately apparent: exploration decision-makers demand resource quantifications and traceable error metrics, not generative hallucinations of questionable provenance. Definitely an area to keep an eye on in 2026!

Research Corner

In the academic world, there were quite a few publications this year that stood out to me:

Despite the amazing progress in the research world, I don’t think we’ve had our “Attention Is All You Need” moment in AI subsurface modeling yet. The closest parallel we have was back in 1987: Occam’s Inversion by the Constables and Bob Parker, the landmark paper on regularized inverse theory. Perhaps 2026 will house the breakthrough in research we need? Maybe it’ll be a NeRF, a Geospatial Reasoning Model, or something we haven’t even named yet!

My personal 2025

2025 was a big year of growth for me. I’m still at Terra AI, doing cool geophysics and subsurface inference work (no updates in this blog post, but hopefully I’ll be able to talk about the work in the future). This year I leaned heavily into using LLMs, both as coding agents as well as API tools, and ended up winning the “Best use of API” the official Meta Llamacon hackathon, using their Llama 4 models to turn geological documents into 3D models! I also presented at the CMU Workshop on Neural Simulation-Based Inference. It was wonderful meeting so many talented statisticians, and telling them all about the compelling problems we have in geosciences. Here’s to another year of exciting tech releases and advances in geosciences!

Cover Photo by Sebastian Pichler on Unsplash

Share this:

Related

Published by posgeo

My name is William Davis. I am a geophysics and data researcher at Scripps Institution of Oceanography. This is a blog about projects I am working on.
View all posts by posgeo

Post navigation

Leave a comment Cancel reply

Δ
/* <![CDATA[ */ document.getElementById( "ak_js_1" ).setAttribute( "value", ( new Date() ).getTime() ); /* ]]> */

Who am I?

My name is William Davis, I’m a senior computational scientist at Terra AI. Previously I was a Miles postdoctoral fellow at the Institute of Geophysics and Planetary Physics at UC San Diego.

Archives

Search

© 2026 PosGeo

Create a free website or blog at WordPress.com.