Physics-informed machine learning for fatigue life prediction of Cf/SiC ceramic matrix composites

Abstract

C_f/SiC ceramic matrix composites are promising materials for high-temperature load-bearing structures. Their fatigue life reflects both mechanical cycling and environmental degradation. This study proposes a Walker-Basquin physics-informed machine learning framework (WB-PIML) to describe their competing contributions and incorporate life information from tests without observed failure. The framework introduces temperature, environment and exposure time per cycle into the Walker-Basquin relation. Runout records contribute condition-specific life lower bounds through a right-censored likelihood. Environment-shared residual learning captures unexplained life differences while preserving the direction of environmental effects prescribed by the physical trunk. Using 224 literature records (158 fractures and 66 runouts), campaign-disjoint validation yields root mean square error ($\mathrm{RMSE}$), mean absolute error ($\mathrm{MAE}$) and $R^2$ of 1.046, 0.794 and 0.308, respectively. The 95% confidence interval for $R^2$ is [0.019, 0.475], and the weighted fraction of fracture predictions within a factor of five is 54.8%. Relative to residual models with simple Walker-Basquin and Basquin trunks, the full censored negative log-likelihood decreases by 0.122 and 0.131, respectively, with both paired 95% confidence intervals above zero. Relative to the machine learning ensemble (ML-Ens), $\mathrm{RMSE}$ decreases by 0.122 (paired 95% confidence interval: [0.010, 0.234]), but the full score exceeds that of the competing-damage trunk alone by 0.026. The framework unifies mechanical-environmental relations with fracture and runout information for C_f/SiC fatigue life prediction. Engineering applicability requires independent validation given the limited number of campaigns and environmental coverage.