Trustworthiness of machine learning models in manufacturing applications using the example of electronics manufacturing processes

The design phase of products decides about a large portion of their costs. Both applied manufacturing technologies and systems are ultimately defined by the products that have to be assembled. Due to global trends towards e-mobility and regenerative energy, electronic printed circuit boards with high copper content and mixed SMT/THT assembly is the predominant assembly scenario. This leads to thermally challenging soldering operations in THT-soldering as high copper masses dissipate the required heat and no reliable manufacturability check is available. Consequently, in electronic manufacturing complex thermal processes have still to be defined in iterative experimental steps. Reliable models have great potential to reduce time-to-market and scrap during new product introduction. Future manufacturing systems require infrastructure that allows manufacturability checks in the early design phase and automatically provides process programs on the shop floor once the design is finished. However, the lack of availability, long-term experience and consequently trust in automatically, by artificial intelligence generated and optimized process programs, blocks the realization and hinders further productivity gain. In this paper, an ML audit methodology is suggested that mitigates this lack of trust by providing physically plausible prediction series and quantitative risk assessment to the user. This complements typical evaluation scores and adds the model behavior in the process context and transmits a sense of transparency. A neural network is trained with an augmented dataset, exemplary for the prediction of the hole fill in mini wave soldering process. The model is then used to predict series of predictions for increasing soldering time and temperatures comparable to design of experiments. The resulting time versus hole fill plots helps the operator understanding the behavior of the model and thus, demonstrate the potential benefit of this approach to the entire manufacturing site.