A physics-informed neural network approach to fatigue life prediction using small quantity of samples

A physics-informed neural network (PINN) is proposed for fatigue life prediction with small amount of experimental data enhanced by physical models describing the fatigue behavior of materials. A multi-fidelity network architecture is constructed to accommodate the mixed data with different fidelities by embedding the physical models into the hidden neuron as the activation functions. Experimental data of two metallic materials is collected for the validation. The results show that the proposed PINN produced physically consistent and accurate results, and performed well in the extrapolative fatigue life prediction.