Can a language model help protect spacecraft polymers from atomic oxygen?

Low Earth orbit is not a gentle place for spacecraft materials. Atomic oxygen—individual oxygen atoms moving at high speeds relative to a satellite—can react with and wear away exposed polymers. Solar radiation, charged particles and temperature changes also affect materials in orbit. Damage to exterior films can compromise spacecraft systems and shorten a mission’s useful life.

Finding better materials is difficult because reliable in-space erosion measurements are expensive to obtain. Ground tests offer a more accessible alternative, but laboratory conditions do not fully reproduce the speed, direction and surrounding environment of oxygen in orbit. Could a machine-learning model help researchers use existing spaceflight measurements to narrow down which polymers are worth testing next?

Learning from past missions

Jared Solovei and colleagues trained a version of OpenAI’s GPT-4o on 201 measurements gathered from NASA exposure missions, including MISSE, LDEF, EOIM III and STS-8. The records cover 55 polymer chemistries, with repeated measurements for some materials under different conditions. The model’s task was to predict erosion yield: the volume of material lost per incident oxygen atom. A lower yield indicates less erosion per incoming atom.

The researchers supplied the model with polymer information—including a text description and a compact representation of its chemical repeat unit—along with exposure details such as atomic-oxygen fluence, solar exposure, spacecraft orientation and mission duration. They transformed erosion yields to a logarithmic scale, which is useful when measurements span many orders of magnitude. The fine-tuning examples paired these inputs with measured yields.

A key choice was how to divide the data for testing. In a random split, measurements of the same polymer chemistry could appear in both the training and test sets. That tests predictions for familiar chemistries under different conditions, but may make performance look better if the model has already encountered that chemistry. To examine performance on unfamiliar materials, the team also used a restricted-group split: test chemistries were excluded from training. Only 29 measurements met the criteria for this more demanding test, and all were from samples facing the direction of travel, known as the Ram orientation.

The team compared GPT-4o with Gaussian process regression, a conventional machine-learning method. They also tested whether adding film thickness and the number of layers improved predictions.

Promising results, with boundaries

On the restricted-group test, GPT-4o achieved an order-of-magnitude error of 0.154 and a log-scaled coefficient of determination of 0.719. In plain terms, these metrics describe prediction error and how well predictions account for variation in measured yields on a logarithmic scale. The authors report that the language model outperformed the Gaussian-process benchmark on this split.

That advantage did not hold for every test. On the random split, Gaussian process regression performed better overall, with the difference attributed to its predictions in the lowest-fluence group. The authors note that this low-fluence regime may be less relevant to screening materials intended to withstand substantial exposure. The random test also included extreme yield values poorly represented in training, while the restricted-group test covered a narrower range. So the apparently stronger score on the chemically disjoint test does not mean that test was universally more difficult: the two splits tested different situations and value ranges.

Adding thickness or layer count did not clearly improve performance in the researchers’ feature tests. This result suggests these measurements contributed little predictive value in this dataset. It does not establish that thickness and layering never matter physically, or that polymer architecture can be ignored in designing real spacecraft materials.

A screening tool, not a substitute for flight data

The study’s central promise is practical: a model might help prioritize candidate polymers before expensive space exposure experiments. It can also accept some information as ordinary text, avoiding the extensive numerical feature engineering required by the comparison model. But the evidence remains limited. The dataset has only 201 measurements, is concentrated in a few chemistries—Kapton and FEP account for more than a third of entries—and excludes some samples whose structures could not be represented consistently. Simplifications in how copolymers and polymer structures were encoded may also affect what the model learned.

Most importantly, 29 measurements from Ram-facing samples provide a narrow basis for claims about predicting unseen chemistry. The authors themselves caution that predictions for new or underrepresented polymers need care. A model score on held-out data is not proof that its recommendations will work in a future spacecraft, where materials encounter complex combinations of environmental stresses.

This is a preprint: the supplied source does not establish that the work has undergone peer review. Its reported results should therefore be treated as provisional. The authors suggest that more chemically diverse flight measurements, alongside ground experiments and simulations that can be connected reliably to flight data, could strengthen future models. For now, the work presents an encouraging way to organize and learn from scarce measurements—not a replacement for validating materials in space.