Physics‐Guided Neural Network Modeling and Process–Performance Mapping of Carbon Fiber/Polyphenylene Sulfide Composites Under Repeated Thermal Histories

ABSTRACT

Carbon fiber/polyphenylene sulfide (CF/PPS) components can experience repeated thermal histories during manufacturing, repair, and reprocessing, yet their cumulative process–structure–property response remains insufficiently understood. Laminates were subjected to five thermal cycles at 300°C, 340°C, 350°C, and 400°C and characterized by differential scanning calorimetry (DSC), attenuated total reflection Fourier‐transform infrared spectroscopy (ATR‐FTIR), and short‐beam testing. After 5 cycles, crystallinity decreased by 2.1% at 300°C and 69.4% at 400°C, where the melting peak temperature decreased by 17.7°C. The oxidation index showed no monotonic trend, although ATR‐FTIR does not exclude subsurface or bulk degradation. Short‐beam strength (SBS) varied non‐monotonically, remaining comparatively high at 340°C–350°C despite decreasing crystallinity and deteriorating markedly at 400°C. A physics‐guided neural network (PGNN) incorporating DSC‐derived descriptors and phenomenological constraints was evaluated by condition‐level five‐fold cross‐validation across 20 conditions. A 2 × 8 architecture yielded out‐of‐fold R ² values of 0.77, 0.91, and 0.79 for crystallinity, melting peak temperature shift, and SBS, respectively, with an SBS RMSE of 9.2 MPa. The resulting process–performance map distinguished favorable, transitional, and unfavorable temperature–cycle combinations within the investigated domain and showed progressive contraction of the favorable region with accumulated cycling, enabling comparative screening of repeated thermal‐processing conditions.