🔬 materials science

Data-driven prediction of atomic-oxygen erosion yield in polymers from bulk physical properties

This study evaluates a data-driven regression ensemble for predicting atomic-oxygen erosion yields in polymers using bulk physical properties, demonstrating moderate success on historical MISSE-2 data and limited discrimination on new polymers due to narrow target ranges, while highlighting the need for reconciled measurement records to validate ground-based screening models.

Original authors: Youngwoo Choi, Doeun Koo, Seungbum Hong, Hyun Jung Kim

Published 2026-10-07

📖 6 min read🧠 Deep dive

CC0 1.0

Original authors: Youngwoo Choi, Doeun Koo, Seungbum Hong, Hyun Jung Kim

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

High above the Earth, in the thin layer of atmosphere known as low Earth orbit, spacecraft are constantly bombarded by a silent, invisible force: individual oxygen atoms. Unlike the oxygen we breathe, which exists as pairs of molecules, these solitary atoms travel at tremendous speeds, striking anything they touch with the energy of a tiny, hyper-fast bullet. For the organic polymers used to build satellites, space stations, and solar panels, this bombardment is the most aggressive form of wear and tear they face. When these atoms hit a plastic surface, they break chemical bonds and rip pieces of the material away, causing it to lose mass, become rough, and eventually fail. This process, known as atomic oxygen erosion, can destroy thermal blankets, degrade solar arrays, and compromise the structural integrity of a mission. For decades, engineers have relied on a slow and expensive method to figure out which plastics can survive this environment: they send physical samples into space, wait years for them to return, and then measure how much material was lost. This cycle is too slow for the rapid pace of modern space exploration, leaving designers with few tools to predict how a new material will hold up before they ever launch it.

To solve this problem, a team of researchers at the Korea Advanced Institute of Science and Technology developed a new way to predict this erosion using only the basic physical properties of the plastic. Instead of waiting for a spaceflight, they trained a computer model on a collection of data from 37 different polymers that had already been tested in space. They fed the computer information that is easy to find in standard engineering handbooks, such as how dense the material is, how much heat it takes to melt it, how much it expands when heated, and the ratio of different atoms like carbon, hydrogen, and oxygen within its chemical makeup. The goal was to see if these simple, everyday numbers could tell the computer how quickly a specific plastic would be eaten away by the atomic oxygen in space. The researchers built a sophisticated system that combined the predictions of several different mathematical models, creating a tool that could instantly estimate the erosion rate for any new polymer, provided its basic properties were known.

When the team tested this tool on nine new polymers that had been exposed to space after the original data was collected, the results were promising but modest. The model successfully predicted the erosion rates for all nine materials within a factor of two, meaning the predicted value was never more than double or less than half of the actual measured value. This level of accuracy is significant because it places the prediction in the correct ballpark, allowing engineers to quickly screen out materials that are likely to fail. However, the researchers were careful to note that the range of erosion rates in these new samples was very narrow, and a simple average of the old data performed almost as well as their complex model. This suggests that while the tool is useful for a first pass, it is not yet a perfect crystal ball that can replace the need for real-world testing. The study highlights that predicting space durability is difficult when the materials being tested are chemically similar to those already in the database.

To push the model further, the researchers applied it to a set of four different grades of a commercial polyimide film, a type of plastic often used in space applications. They compared the computer's predictions against new experiments conducted in a ground-based laboratory that simulates the atomic oxygen environment. The model correctly identified the ranking of the four grades: it predicted that one specific grade would erode the fastest and another the slowest, and the ground-based experiments confirmed this exact order. This success gave the team confidence that the model was capturing real physical relationships rather than just guessing. They also examined the surfaces of these materials after exposure using powerful microscopes and chemical scanners. They found that the grades which resisted erosion better tended to leave behind a slightly higher amount of solid residue after being heated, and their surfaces showed different patterns of wear. This observation supports the idea that materials which form a protective layer of inorganic residue as they erode might last longer, a concept that the model's inputs helped to identify.

The researchers also tested the model on a "mystery" polymer film whose chemical composition was kept secret from them. They were given only the basic physical properties and a measurement of how much residue the material left behind after heating. The model predicted an erosion rate that was within a factor of two of the actual measurement taken in the laboratory. This was a crucial test because it showed the model could handle a material it had never seen before, without knowing its specific chemical formula. However, the researchers noted that even for this mystery sample, a simple average of the training data performed nearly as well as their advanced model. This finding serves as a reminder that while the new tool is a step forward, the small size of the available space data makes it hard to prove that the complex model is vastly superior to simpler methods.

Ultimately, this work demonstrates that it is possible to build a rapid screening tool for space materials using only standard physical data. The model provides a way to estimate how long a polymer might last in orbit without waiting years for a spaceflight experiment. The researchers emphasize that this tool should be viewed as a screening framework to help engineers narrow down their choices, rather than a definitive replacement for physical testing. The study also clarifies that while certain properties like density and residue formation appear to be important, the exact mechanism of how these properties protect the material is still not fully understood. By combining data from space, ground experiments, and surface analysis, the team has laid the groundwork for a more robust system. Future work will need to include a wider variety of materials and more precise measurements to turn this screening tool into a universally reliable guide for designing the spacecraft of tomorrow.

Technical Summary: Data-driven prediction of atomic-oxygen erosion yield in polymers from bulk physical properties

Problem Statement
Atomic oxygen (AO) is the most aggressive surface-degradation mechanism for organic polymers in Low Earth Orbit (LEO), causing mass loss, surface roughening, and mechanical failure. While the Materials International Space Station Experiment (MISSE-2) provided a benchmark dataset of erosion yields ( $E_y$ ) for 41 polymers, generating new experimental data is slow, expensive, and limited by the scarcity of flight-validated materials. Existing predictive tools, such as those by Banks et al., rely on hand-encoded chemical structure descriptors and bond inventories, which are difficult to automate for high-throughput screening. Furthermore, ground-based AO simulators often yield results that differ by factors of 2–5, creating uncertainty in material selection for new spacecraft missions. There is a lack of a data-driven model that utilizes only widely tabulated bulk physical properties and elemental composition ratios to predict $E_y$ without requiring manual structural encoding.

Methodology
The authors developed a regression ensemble model trained on the 37 polymers from the MISSE-2 PEACE dataset (excluding PVC due to lower-bound reporting). The model utilizes 13 input features:

Eight bulk physical properties: Density, ash mass fraction, oxygen index, glass transition temperature ( $T_g$ ), total-mass-loss outgassing, linear coefficient of thermal expansion (CTE), ultimate tensile strength, and specific heat capacity.
Five elemental-composition ratios: H/C, O/C, N/C, F/C, and other-heteroatom/C.

The ensemble consists of four base learners: RidgeCV, a Gaussian Process Regressor, an XGBoost Random Forest-style regressor, and an XGBoost sequential gradient booster. Each learner predicts the log-transformed target ( $\log_{10} E_y$ ), and the final prediction is the geometric combination of these four outputs with equal weights.

Crucially, the model development followed a specific protocol: feature selection (top eight physical features) and ensemble scheme selection were performed on the complete training table before leave-one-out (LOO) splits. Consequently, the authors note that the archived out-of-fold (OOF) scores are conditional on these selection steps and do not represent an unbiased validation of the entire model-development procedure.

Key Results

Internal Validation (MISSE-2 Training Set): The archived ensemble achieved an OOF $R^2(\log)$ of 0.768. 30 out of 37 samples were predicted within a factor of two of the measured values. However, the authors highlight that the training set has a narrow target range, and simple arithmetic-mean baselines also met the factor-of-two tolerance for many samples, limiting the discrimination power of the model on this specific dataset.
External Orbital Validation (Post-MISSE-2): The model was tested on nine independent polymers not included in the training set. The ensemble achieved an $R^2(\log)$ of −0.037, with all nine predictions falling within a factor of two of the measured values. The training arithmetic mean baseline performed similarly (all within a factor of two), suggesting that the narrow span of the external target range (2.56-fold) makes it difficult to demonstrate broad chemical transferability or model superiority over simple baselines.
Ground-Based Comparison (PLAVIS Grades): The model was applied to four grades of a commercial polyimide (PLAVIS: HM, S1, ESD, N). The model successfully reproduced the experimentally observed ranking of erosion yields ( $E_y$ $E_{y}$ N > ESD > S1 > HM), which corresponds to an AO resistance ranking of HM > S1 > ESD > N.
- Correlation with Residue: The model's performance aligned with Air-thermogravimetric analysis (Air-TGA) residue data. Grades with higher Air-TGA residues (HM, S1) exhibited higher AO resistance. While this supports the model's reliance on ash-related descriptors, the authors caution that Air-TGA residue was used as a proxy input for the ash descriptor, meaning this correlation serves as an internal consistency check rather than independent validation of the ash mechanism.
- Surface Characterization: SEM and XPS analysis revealed that grades with higher resistance (HM, S1) had more compact surface morphologies and detectable Ca/Si surface species, whereas lower-resistance grades (N) showed pronounced surface texturing. However, the authors state these observations do not definitively establish a protective-layer mechanism or identify specific shielding compounds.
Mystery Sample Case Study: A polymer with undisclosed composition was evaluated. The model predicted an $E_y$ within a factor of 1.92 of the ground-based measurement. While the ensemble prediction was closer to the measurement than the training arithmetic mean, the training geometric-mean baseline yielded a similar fold-error (1.96), again limiting claims of ensemble superiority for this specific case.

Significance and Claims
The paper positions the work as a screening framework rather than a universally accurate predictor. The primary contributions are:

Rapid Pre-screening: The model demonstrates that a surrogate trained on tabulated bulk properties and elemental ratios can provide rapid (millisecond) estimates of AO erosion yield, compressing the flight-experiment cycle for initial material down-selection.
Design Guidance: The feature importance analysis (highlighting glass transition temperature, F/C ratio, oxygen index, and ash fraction) offers interpretable design guidance for polymer chemists, indicating which experimentally accessible properties modulate AO durability.
Validation of Proxy Inputs: The study validates the use of Air-TGA residue as a proxy for RF-plasma ash in predictive models, showing consistent grade-dependent trends with experimental AO resistance, though it stops short of proving a causal protective mechanism.

Limitations and Future Directions
The authors explicitly acknowledge several limitations:

Small and Narrow Dataset: The training set is small (n=37) and chemically related; the external datasets span narrow erosion-yield ranges where simple baselines remain competitive.
Conditional Validation: The feature selection and model architecture were not nested within every cross-validation split, meaning the reported scores are conditional on the full training table.
Lack of Independent Mechanism: The study does not establish a definitive causal link between specific features (like ash content) and the erosion mechanism, nor does it prove that the observed correlations apply universally across chemically diverse polymer families.

The paper concludes that future work should focus on prospectively fixed descriptor records, larger and more chemically diverse datasets, and matched pre- and post-AO surface characterization to transform the current screening framework into a more robust and physically interpretable tool.

Technical Summary: Data-driven prediction of atomic-oxygen erosion yield in polymers from bulk physical properties

More like this