Researchers fine-tune GPT-4o to predict atomic-oxygen erosion of spacecraft polymers from 201 NASA measurements

A preprint posted to arXiv on 22 October 2026 by Jared Solovei describes a fine-tuned version of OpenAI's GPT-4o built to estimate how quickly atomic oxygen erodes polymers on satellites in low Earth orbit. The model was trained on 201 erosion-yield measurements compiled from multiple NASA exposure missions, not on a new flight campaign. Atomic oxygen, a reactive form of oxygen common a few hundred kilometers up, is a chief reason thin polymer films, coatings, and thermal blankets wear out in orbit. The paper has not been peer-reviewed.

The preprint reports an order-of-magnitude error of 0.154 and a log-scaled coefficient of determination of 0.719. The typical miss was a fraction of an order of magnitude, and the model tracked a substantial share of the variation once yields were compared on a log scale. The authors say it beat conventional machine-learning methods without heavy feature engineering. Thickness and layering did not help, which they read as a sign that chemistry and exposure conditions matter more than bulk structure. Methods note: GPT-4o fine-tuned on 201 historical NASA points, preprint only, with no new flight validation shown.

The stake is how long a satellite stays useful. Most active spacecraft already fly in low Earth orbit, and operators keep adding more for broadband and imaging. Atomic oxygen thins polymer films on solar arrays, insulation, and composites, so a weaker material means earlier replacement and more objects to track. The preprint cites a projected debris cascade, often called Kessler syndrome, as the reason to stretch those lifetimes; that cascade is motivation, not a result of this work. The testable claim is a chemistry-first screen that might cut how many polymers need a flight exposure.

What would strengthen the evidence is a test on polymers the model never saw, then a comparison with a new low-Earth-orbit exposure rather than another pass over the same 201 NASA points. Peer review should also require the train-test split, the prompt design, and a check that mission data did not leak into the fine-tune. If those hold, the near-term use is screening, not certification: a short list of chemistries for the next materials flight experiment. An error of 0.154 on historical data is not a reason to skip ground testing.