Geometries of ceramic parts for high-temperature sealing have great influence on their compression-resilience behaviors. In this work, an accurate and large-scale artificial neural network (ANN) was established to match the relationship between structural parameters and mechanical properties of ZrO 2 parts fabricated by 3D printing. Four geometry parameters of the designed ZrO 2 parts were imported as input and apparent Young's modulus and maximum deformation simulated by finite element method (FEM) were imported as output. FEM calculation provided 400 groups of data for the training of ANN, which greatly improved the predicted accuracy of the network. The predicted results show the mechanical performance of the parts with a range of modulus from 9.24 × 10 −3 GPa to 100.35 × 10 −3 GPa and a range of maximum deformation from 2.32% to 5.80% can be forecasted with error less than 8%. Based on the optimized structural parameters, the designed ZrO 2 parts were fabricated by Direct Ink Writing (DIW) technique. The experimental compressionrebound property is comparable to that of ANN prediction. It demonstrates that the combined method of ANN and FEM is a preferable way to optimize the structure and guide the fabrication of complex ceramic parts by 3D printing method.
K E Y W O R D Sadditive manufacturing, artificial neural network, finite element analysis, sealing 218 | FAN et Al. where N is the number of output, y n is the predicted value of the n th item, d n is the target value of the n th , and Q is the number of training cycle. Then the 400 groups of data were divided randomly into three parts: training groups, validation groups, and test groups with the ratio of 70%, 15%, and 15%. The network would stop training when the MSE of validation did not decrease continuously more than six times, otherwise the network was not good at generalization. Similarly, the network would also stop training when the training epochs came to 100 or the value of MSE was less than 0.00001.