Machine learning based prediction of mechanical properties in carbon fiber recovered through pyrolysis

The limitations of traditional pyrolysis technologies of recovered carbon fiber included long processing time, low efficiency, and unclear links between process parameters and performance. Moreover, pyrolysis alone formed residual coke on carbon fiber surfaces, severely impairing reusability. Therefore, this study adopted a combination of pyrolysis oxidation recycling to obtain clean carbon fibers, and introduced an innovative machine learning optimization framework to predict and analyze key recycling parameters. Specifically, a database of raw material characteristics and process parameters was built via systematic literature research. Subsequently, four machine learning models were integrated to predict mechanical performance, and the random forest model performed the best with determination coefficients of 0.9608 and 0.9419, respectively. Additionally, partial dependency graph analysis quantified the process window and determined the optimal parameters. Optimal tensile modulus was attained with a 5 ℃/min pyrolysis rate, 570 ℃ oxidation temperature, and 45%-75% carbon fiber mass fraction. For tensile strength, the best parameters were 557 ℃ oxidation temperature and 45 mins oxidation time. A 5 ℃/min pyrolysis rate stabilized tensile modulus, while 500 ℃ preserved tensile strength. Additionally, 75 wt% carbon content combined with high fiber mass fraction enhanced mechanical performance. The foundation was laid for subsequent pyrolysis‑oxidation experiments on carbon‑fiber recovery, and the recovery efficiency was improved. Overall, this study can provide more accurate theoretical support for carbon fiber recycling technology.