Research report: Unsupervised learning-based analysis of degradation behavior in filler-containing amine-cured epoxy formulations

Adhesives quietly hold together the modern world, from semiconductor packages to the lightweight joints that make cars and aircraft lighter. But like every material, they age, and the chemistry of that aging has been notoriously hard to untangle. A new study published in Results in Chemistry now shows that an unsupervised machine-learning method, principal component analysis (PCA), can systematically organize the complex spectral fingerprints of degrading epoxy adhesives and link them to the loss of mechanical strength, with results that differ strikingly depending on which curing agent was used.

The curing agents were triethylenetetramine (TETA), an aliphatic polyamine; p,p’-methylenedianiline (MDA), an aromatic diamine; and POLYMIDE L-4051 (AMAM), a commercial amidoamine. All three were mixed into a common base resin supplied by Cemedine Co., Ltd., containing epoxy resin, a CTBN-modified epoxy, silica, calcium carbonate, and calcium oxide. The fillers were included to adjust viscosity, and the calcium carbonate in particular would later prove to leave its own spectral signature.

Specimens were cast as 0.3-millimeter-thick sheets between aluminum plates and cured under schedules tailored to each agent: TETA at 100 °C for one hour, MDA at 160 °C for one hour, and AMAM at 100 °C for 30 minutes. The cured sheets were then punched into dumbbell-shaped tensile specimens. Accelerated aging was carried out under three conditions: water immersion at 63 °C and 87 °C, and exposure to humid air at 120 °C with 100 percent relative humidity. Immersion tests ran through fifteen predefined intervals, from one hour up to a full year, with three specimens tested at each condition. The severe 120 °C condition was stopped after one week because degradation was so rapid.

After each aging period, the team measured bulk tensile strength and recorded infrared spectra with a handheld FTIR spectrometer using the attenuated total reflection method, at four locations on each fractured specimen. The spectra were deliberately left unprocessed apart from a common offset adjustment, so that the analysis would capture the raw variations before any model-dependent assumptions were introduced. The researchers caution that because specimens were measured shortly after removal from wet conditions, residual surface moisture could not be distinguished from water absorbed within the resin, so the OH stretching region was interpreted only qualitatively as a moisture-related signal.

PCA was then applied independently to the spectra from each degradation environment, with mean centering but no variance scaling, since the absolute magnitude of spectral changes was considered physically meaningful. The first three principal components generally captured around 90 percent of the variance. Across the systems, the dominant components consistently mapped onto three spectral themes: variations in the OH stretching region at 3000 to 3600 per centimeter, interpreted as moisture-related; changes near 1700 and 1100 per centimeter involving carbonyl, CO, and CN vibrations, consistent with oxidation-associated processes; and features near 700, 900, and 1400 per centimeter where calcium carbonate bands appear, pointing to the filler phase. The authors stress that these labels describe dominant spectral features and working hypotheses, not experimentally verified degradation mechanisms.

Bootstrap resampling with 100 iterations confirmed that the loading patterns were robust, with mean Pearson correlation coefficients between original and resampled loading vectors ranging from 0.923 to 0.999 for the first three components. Yet the degree of separation among the components depended strongly on the curing agent. For TETA at 63 °C and AMAM at 63 °C, the three factors were clearly distinguishable. For MDA at 63 °C, by contrast, the water-related and oxidation-related variations evolved on similar timescales and appeared strongly superimposed, making them hard to separate by temporal behavior alone. At 87 °C, the MDA system’s factors did separate, suggesting that with increasing temperature the individual degradation processes diverged in rate and dominance.

The AMAM system behaved in the most surprising way. Its tensile strength did not decline monotonically: it first dropped, then partially recovered, and finally fell again. The initial drop, seen in all three systems, is consistent with plasticization by early water ingress, which can lower the glass-transition temperature. The later recovery and decline tracked the second principal component, whose loading pattern combined carbonyl-related features with changes in the hydrogen-bonding environment around 3300 per centimeter. The authors note that interactions between water and polymer networks can induce network relaxation or rearrangement that temporarily improves mechanical strength, and that amidoamine groups have a high affinity for water. A sharp final-stage strength loss at 87 °C instead coincided with a sudden change in the third component, the calcium carbonate-associated factor, hinting that filler-related or interfacial changes may become important late in degradation, though the authors emphasize this remains a tentative hypothesis without independent experimental confirmation.

One of the most striking demonstrations of the method’s sensitivity came from an anomaly that would have been easy to miss. The 168-hour observations in the AMAM system deviated markedly from the temporal trajectories of the second component at both immersion temperatures. Statistical analysis using externally Studentized residuals from a quadratic regression identified these points as outliers, with residuals of 5.43 and −6.93 at 63 and 87 °C respectively, while all other aging times fell below 1.87 in absolute value. Inspection of the corresponding spectra revealed atypical baseline behavior in broad regions, including 1700 to 2100 and 2250 to 2600 per centimeter, suggesting a measurement-related influence. Excluding these points changed the correlation between the second component and tensile strength at 63 °C from r = −0.562 to r = −0.856, and at 87 °C from r = 0.013 to r = −0.751, revealing a moderate-to-strong stage-dependent relationship that global correlations had obscured.

Indeed, the global correlation analyses across all aging periods showed that no single principal component was consistently associated with tensile-strength evolution across all curing systems and environments. For the MDA system at 87 °C, the water-related first component correlated strongly with strength (r = −0.861), while the oxidation-related third component, though accounting for roughly ten percent of the variance, showed a much weaker relationship, a pattern the authors suggest may reflect the protective influence of the rigid aromatic network. The authors also note that because FTIR measurements were performed on fractured specimens after tensile testing, the spectra may contain contributions from specimen damage itself, and they call for future measurements on independent, untested specimens.

The study’s authors are careful to bound their claims. All formulations shared a common CTBN-modified, filler-containing base resin, so the comparisons primarily reflect differences in curing-agent chemistry and the resulting network structures; the influence of CTBN, crosslink density, and glass-transition temperature was not independently evaluated. The tests used bulk tensile specimens rather than bonded joints, so the measured strengths reflect the intrinsic properties of the cured adhesive, not the durability of an adhesive joint. Even so, the framework offers something genuinely useful: a way to identify which spectral factors matter at which stages of degradation, to flag subtle measurement anomalies invisible to routine inspection, and to guide the design of the next round of experiments. As the authors conclude, the approach is not merely a spectral analysis technique but an integrated method for accelerating the hypothesis-generation and verification cycle in material degradation research, with potential value for lifetime assessment and reliability engineering in the industries that depend on adhesives every day.

Subject of Research: Machine-learning analysis of infrared spectra to track degradation in amine-cured epoxy adhesives

Article Title: Unsupervised learning-based analysis of degradation behavior in filler-containing amine-cured epoxy formulations

Article References: Yamaura, D., Shimamoto, K., Terasaki, N., & Akiyama, H. (2026). Unsupervised learning-based analysis of degradation behavior in filler-containing amine-cured epoxy formulations. Results in Chemistry, 32, Article 104025. https://doi.org/10.1016/j.rechem.2026.104025

Image Credits: AI Generated

DOI: 10.1016/j.rechem.2026.104025

Keywords: epoxy adhesives, principal component analysis, FTIR spectroscopy, degradation, curing agents, tensile strength, moisture uptake, oxidation, calcium carbonate filler, machine learning, materials reliability, hydrothermal aging