Process Data Analysis for Improved Burn-In Strategies Based on Complementary AI Models

Burn-in (BI) tests are the industry standard to screen out the early life failures of semiconductors. Advanced sampling and test strategies allow to reduce BI times or sample size without affecting the defined quality targets. A new BI approach introduces a lot-specific health factor h that correlates to the probability of early failures. In our approach, the health factor of a specific wafer lot is derived from the Advanced Process Control (APC) system that logs all meta and logistics data of the production process, but not the raw sensor data. Complementary AI models were investigated to provide health indicators with a high correlation to early failures.

A practical study has been performed based on real APC data and several known BI defects. Due to the amount of data, big data strategies had to be applied and tested to reduce the computational demands. Our investigation shows that a combination of a binary classifier and an LSTM autoencoder model allow for a good assessment of the health factor. In addition, the autoencoder allows to identify and visualize potential issues of the production process via its loss function. That enables process engineers to assess and to investigate potential risks and issues of a specific lot.

Keywords: Semiconductor devices, Burn-in tests, Early life failure rate, Advanced process control, Deep neural networks.

Download PDF