Generative adversarial networks enable outlier detection and property monitoring for additive manufacturing of complex structures

Although additive manufacturing processes have matured in many areas, difficulties with regard to printing accuracy persist. Possible defects are, e.g., the generation of unwanted internal pores or a lack of fusion between layers. In general, defects result in a deviation between as-planned and as-built geometries. Assessing the influence of such deviations on effective (mechanical) properties requires either experimental testing or numerical investigation, both of which are often too complex for being used in time-critical production settings. Previous work with the Finite Cell Method (FCM) has shown that image-based computational homogenization can assist in principle in quality monitoring of produced parts by providing effective mechanical properties of as-built geometries. However, conducting FCM computations online for each and every part of a series production demands a prohibitive amount of computational resources. The paper at hand suggests a remedy to this problem by using generative adversarial networks. This manuscript presents an integrated computational workflow for outlier detection and property monitoring, which in principle consists of four steps: (i) Carry out experiments under varying process conditions and collect respective (micro)-structural images. (ii) Compute effective properties for these images using FCM. (iii) Compose a generative adversarial network (GAN) training dataset from those images that meet desired effective properties. (iv) Use the GAN discriminator to detect outliers in series production. To this end, the discriminator is used as a classifier on as-built parts to judge whether an as-built structure is acceptable or defective. The viability of the approach is demonstrated on additively manufactured lattice structures whose geometry is acquired after production via computed tomography. The methodology is not only applicable for automated property monitoring but potentially also for reliability estimates of neural network-based property predictors.