Machine Learning for Celestial Body Analysis: How We Use AI in Astronomy Research

When asteroids teach you to build better AI

This article is different from the others in this series. It’s about my scientific research, not our client work. But I’m including it because it illustrates something I believe strongly: that genuine scientific practice produces better AI engineers than any bootcamp or certification.

My PhD research focused on a specific problem in celestial mechanics: identifying asteroids trapped in mean-motion resonances with planets. A mean-motion resonance occurs when two orbiting bodies exert regular, periodic gravitational influence on each other — for example, an asteroid completing exactly three orbits for every two orbits of Jupiter. These resonances can stabilise or destabilise asteroid orbits over millions of years, and identifying which asteroids are in resonance (and which type) matters for understanding the long-term evolution of the solar system.

The problem

The dataset is the entire catalogue of known asteroids — hundreds of thousands of objects, each with orbital elements that evolve over time. Determining whether an asteroid is in a resonance requires numerical integration of its orbit over millions of years and analysis of the resulting time series for periodic patterns. Doing this manually for each asteroid is computationally expensive and practically impossible at catalogue scale.

The traditional approach uses numerical indicators — computing specific orbital angles and checking whether they oscillate (resonance) or circulate (no resonance). This works but is slow and misses complex resonance types, particularly three-body resonances involving two planets and an asteroid simultaneously.

The ML approach

We developed a machine learning approach to classify asteroid orbits as resonant or non-resonant based on features extracted from their orbital evolution. The pipeline processes orbital integration data, extracts time-series features (frequencies, amplitudes, stability indicators), and uses trained classifiers to identify resonance patterns.

The result was a Python package published in Astronomy and Computing, with the underlying research published in Monthly Notices of the Royal Astronomical Society and Icarus. The package lets any researcher apply our classification methodology to their own asteroid data.

What this has to do with building AI for clients

The connection isn’t abstract — it’s methodological. Three specific lessons from astronomy research directly inform how we build AI systems.

Evaluation methodology matters more than model selection. In astronomy, the ground truth for resonance classification comes from decades of numerical integration and expert analysis. We spent as much time validating our classification results against these ground truths as we did building the classifier. The same discipline applies to legal AI (is the citation actually accurate?), financial AI (does the fraud model catch real fraud?), and educational AI (does the essay grading match expert assessment?).

Feature engineering outweighs model sophistication. The key to our asteroid classifier wasn’t a fancy deep learning architecture — it was choosing the right features to extract from the orbital time series. Fourier analysis, Lyapunov exponents, resonance angle statistics — these domain-specific features made a simple random forest competitive with more complex approaches. In commercial AI, the same principle holds: understanding your domain well enough to choose the right features matters more than choosing the fanciest model.

Reproducibility is non-negotiable. We published our code, our data, and our methodology so that other researchers could reproduce and verify our results. In commercial AI development, this translates to documented evaluation pipelines, versioned model configurations, and audit trails that let you explain why the system makes the decisions it makes.

“Science teaches you to be honest about what you don’t know. Every paper includes a limitations section. Every result comes with error bars. I wish the AI industry adopted this practice more widely. When we deliver an AI system to a client, we include a ‘limitations’ document — here’s what the system does well, here’s where it struggles, here’s the error rate you should expect. Clients appreciate it more than inflated promises.”

The broader point

Having a founder with published scientific research isn’t just a credential. It’s a way of thinking that permeates how the team approaches problems. When a client asks “how accurate is this model?”, we don’t say “very accurate.” We say “retrieval precision is 89% at top-5, answer faithfulness is 94%, and citation accuracy is 97% — here’s the evaluation methodology we used to measure this.” That specificity comes from scientific training, and it’s one of the things that distinguishes our work.

---

Need AI development with scientific rigour? Contact us — we bring the methodology, not just the models.