PSI - Issue 79

Henrik Petersson et al. / Procedia Structural Integrity 79 (2026) 298–305

303

Fig. 2. Experimental data and prediction of fatigue life, comparison ISO model and PINN model with 95% probability interval.

factors. Around the level of 300 MPa, the predicted life would be 34400 cycles for the PINN model and 26800 cycles for the ISO model. This is a di ff erence of 28% and could be utilized to reduce conservatism without increasing the risks of fatigue failure for the material. Consider if there were a lack of experimental data in a certain fatigue life region. Then, the uncertainty is ex pected to increase. This concept is investigated by removing some experimental data, in this case data between 260 MPa and 300 MPa. Thus, the probability interval changes in a way that can require additional estimation of the smoothing parameter in the Gaussian kernel function. Another way to control the probability interval is by introducing the uncertainty factor mentioned previously in Equation 15, which is dependent on the distance from the experiment data, by implementing this the results in Figure 3 are obtained. Through this uncertainty factor, the probability interval is enlarged in the region where the experimental data are removed. Doing this requires the use of a larger safety factor, not to run into fatigue failure. In this paper it has been shown that a PINN model can be trained in such a way that it captures the behaviour of the stress vs fatigue life curve in a satisfactory way. Both the upper and lower stress limits are captured and the experimental data is estimated in a way that can decrease safety factors in a model without increasing the risks of failure. The PINN model with its 95% probability interval captures the experimental data in a su ffi ciently well, and in regions of high certainties even lower nominal safety factors. The current model does not include scatter around the stress limit and the fatigue limit. Di ff erent and more complex kernel functions can in the future be explored and implemented to describe the variance of the probability interval. 4. Conclusions