How KoBold Metals Monitors 1000s of Geoscience and ML Experiments in One Place

KoBold Metals discovers and develops new sources of critical metals like cobalt, lithium, nickel, and copper. Their teams of geologists, data scientists, and engineers run a wide variety of computational and ML experiments on everything from geochemical and remote sensing data to inversions of electromagnetic surveys to make better exploration decisions, faster.

The challenge

Before Neptune, KoBold didn’t have a standard way to track experiments. Individual data scientists were developing pipelines, running simulations, and logging outputs in whatever way they could, often in Jupyter notebooks or local logs. When results were shared internally, important details, like parameters or code used for training, were not formally tracked. 

As the scientific computing team began building more reusable infrastructure, this lack of standardized tracking became a serious obstacle. They needed a way to separate code from execution and systematically track every run—inputs, parameters, results, and artifacts—so they and other team members could understand and reproduce the work.

Reproducibility and traceability are essential to KoBold. Our experiments inform multi-million dollar decisions about where to explore and how to develop what we find. We can’t make those decisions with confidence if we can’t trace the process that led to them.

Josh Bauer Staff Machine Learning Engineer at KoBold Metals

Logging and inspecting experiments at scale

At KoBold, researchers run thousands of experiments per month across varied domains, from surrogate models for Maxwell’s equations to drill core computer vision and mine planning simulations. Many of these experiments are quick iterations or short-lived debug runs, while a smaller number are promising and require deeper inspection.

With Neptune, every experiment is logged automatically from the pipeline. Runs are tagged with metadata (like the location, model type, or exploration phase), parameters are captured from Hydra configs, and outputs, including numerical metrics and interactive artifacts, are uploaded as part of each run.

This allows KoBold researchers to quickly browse past work and focus on what matters:

• Spot unpromising or broken experiments early by checking mid-run logs.
• Tag and search to find the exact configuration that led to an insight.
• Build on past work without rerunning it.
  

One way our team uses Neptune is to monitor experiments as they run. If something is clearly going to take an unreasonable amount of time, we can catch it early and stop it, rather than waiting hours to find out. That visibility mid-run has been really useful.

Liz Main Head of Scientific Computing at KoBold Metals

For our team, Neptune is more like a lab notebook. Most runs aren’t interesting on their own, but for the ones that are, we want a clear history: what configuration we used, what changed, and what the outputs looked like. It helps us stay grounded and iterate methodically.

Josh Bauer Staff Machine Learning Engineer at KoBold Metals

One communication layer between experiments and decisions

Neptune is not just a tracking tool for data scientists. It’s also how exploration decisions are communicated to domain experts and leadership. Researchers include Neptune links or outputs in internal write-ups, Slack messages, and even mine-planning presentations.

Each run contains everything needed to explain a scientific result: inputs and parameters, links to source code, outputs, plots, and interactive figures.

Geologists can explore results directly in the app, zooming into charts and reading parameters without needing to run code. This has been especially important for collaborative decision-making around where to drill, how to plan mine shafts, and how to assess a proposed strategy.

Science can’t just be a black box. Geologists need to see the results, and they need to understand what parameters went into them. Neptune helps us do that.

Liz Main Head of Scientific Computing at KoBold Metals

Well-organized structure, even with many projects and users

Because experiments span many domains—computer vision, physics simulation, geochemistry, optimization—KoBold tracks each modeling effort in its own Neptune project. Some projects are long-running and carefully maintained; others are ephemeral, used for a short burst of analysis and then archived.

Despite the number of projects and runs, the interface remains fast and manageable:

• Each run includes tags and unique identifiers based on location or drill hole index.
• Saved views and search filters help teams quickly surface relevant work.
• Metadata like Git hashes and configuration names are used to link Neptune runs back to code or discussion in GitHub PRs.

This structure supports the reproducibility KoBold values, without forcing unnatural workflows on individual scientists.

The results

Neptune has become Kobold’s system of record for experimental work. The key benefits include:

• Consistent tracking of inputs, outputs, and parameters across 50+ projects.
• Shorter feedback loops by inspecting results mid-run and killing non-promising jobs.
• Cross-functional visibility, as scientists, engineers, and geologists all view the same data.
• Reproducible decision-making for science-driven mine planning.
  

Neptune has been critical for experiment tracking at KoBold. It gives us a transparent, searchable record of our work, something we rely on to do rigorous, applied science. We’re all busy, and if a tool isn’t easy to use, it tends to get skipped. But Neptune strikes the right balance: it’s powerful and easy enough that our team actually uses it.

Liz Main Head of Scientific Computing at KoBold Metals

Thanks to Liz Main and Josh Bauer for helping create this case study!

Product resource

How Navier AI uses Neptune to Rapidly Iterate on Physics Foundation Models

Read more Product resource

Product Demo: An Overview of How to Use Neptune When Training Foundation Models

Read more