How Navier AI uses Neptune to Rapidly Iterate on Physics Foundation Models
Navier AI is building physics foundation models for hardware engineering simulations, with a vision to deliver simulation-quality predictions in seconds rather than days. Their initial focus of computational fluid dynamics (CFD) simulations empower engineering teams, from F1 and aerospace to HVAC and automotive, to make design decisions faster by approximating the results of expensive fluid simulations using deep learning.
Navier trains large-scale models (tens of millions of parameters) on complex, high-fidelity CFD datasets. These models are continuously refined through architecture iterations, dataset variations, and hyperparameter tuning.
The challenge
Initially, Navier AI used Weights & Biases and other logging tools to track experiments but came across challenges with UI slowdowns and platform outages that disrupted their workflow.
As their experiment workloads continued to increase, the Navier team encountered growing issues with robustness, as well as a diverging focus from their needs (shifting away from supporting traditional deep learning training). The tool slowed down progress instead of enabling it.
What Navier needed was a focused, reliable solution purpose-built for experiment tracking. A tool that wouldn’t distract them with unrelated features or slow them down.
Smooth logging that supports iterative workflows
Navier switched their experimentation workflow to Neptune, and it quickly became their primary interface for evaluating and comparing training runs. They gained clarity, speed, and confidence in each iteration.
Each of Navier’s training runs logs over 10 different loss and evaluation metrics for multiple physical fields: field values like velocities and pressures, as well as integrated force coefficients like drag and lift. Some metrics are expensive to compute and are only logged for the best models.
Neptune lets them track all of this in real time without lag or failures. This helps keep GPU hours in check and cycle through architecture tests quickly.
The Navier team was able to fully migrate off of their previous monitoring tool within just a few days.
And the integration with Neptune was straightforward. Navier doesn’t use any framework since their training setup is rather bespoke. But Neptune’s low-level API lets them log exactly what they want, when they want, without adding overhead or complexity.
Faster visual analysis for high-impact decisions
Navier uses Neptune’s reports to centralize all key visualizations per experiment. Oliver tracks training and validation MSE as primary signals, along with per-field physical error metrics, prediction directionality, and physics-based loss function errors. For the best run in a series, he logs additional summary metrics used in customer reporting.
He checks training curves regularly to assess convergence, overfitting, and whether a run should be stopped. Fast response to any chart customization, zoom accuracy, and consistent coloring across sessions make it easy to spot issues or gains.
Those same reports serve as a reference point for summarizing progress to company stakeholders, other team members during team reviews or as a source of truth for engineers deploying the best models to production.
A partner that listens and iterates fast
What sealed the deal for Navier was not just Neptune’s performance. It was the relationship. The team felt heard, supported, and closely aligned with Neptune’s product direction.
“We’re excited to work with a team that actually wants our feedback and acts on it. Every few weeks, we see something new shipped. That’s how we build products, too,” says Oliver.
The results
By switching to Neptune, Navier AI has:
- Eliminated downtime concerns and saved significant time over just a few months;
- Sped up iteration cycles with truly real-time monitoring and insights;
- Built a foundation for further scale;
- Found a partner aligned with their workflow and focus.
Thanks to Oliver Lammas (founding engineer) and Evan Kay (co-founder) for helping create this case study!
Product resource 
