A Methodology for Predictive Life Expectancy of Moisture-Sensitive SMT components using Neural Networks

Abstract

With the increasing demand for climate-neutral production procedures, resource efficiency plays a significant role. In addition, chip shortage happened due to COVID-19, for example, costs for the global automotive industry were $210 billion in revenue in 2021 only, showing the importance of electronic components in the industry and the necessity of taking action to tackle the chip-waste problem in the production line. The circumstances show that chip availability and efficient use are gaining relevance. This also includes the optimal utilization of environmentally sensitive components to avoid waste. As for components sensitive to humidity, moisture is absorbed when they are not handled under dry conditions. In the production, after the opening of sealed packages, this causes problems such as a fracture in micro scales to inflation or popcorning of the component. Established norms like JEDEC often suggest baking the components for de-moisturizing or disposal of the components. The exact environmental condition is hardly taken into account. This yields to extra energy consumption for unnecessary drying on the one hand or waste of components on the other hand.In this contribution, the authors suggest a methodology to predict the weight change of SMT components due to moisture absorption/desorption based on environmental conditions using a Machine Learning model. In this regard, one can decide for each component individually, if it needs to be dried based on the real amount of moisture absorbed and consequently, prevent the component failure in the production procedure. Temperature, humidity, and weight change of components are measured with the Magnetic Suspension Scale (MSC) and Thermogravimetric Analyzer (TGA) over time and under defined environmental conditions. To predict the weight change over a defined period, a Recurrent Neural Network has been trained which has shown a reasonable accuracy in predicting the weight change. Furthermore, to generalize the trained model for different types of components, parameters like dimensions and density are fed into the model to represent their physical characteristics. For the validation three different component types, classified for the Moisture-Sensitivity-Level 4 (MSL) are used.