Learning from Failures in Large Scale Soft Sensing

Scientific publications typically favor studies that show successful results. Things that did not work either are relegated to footnotes, or worse, never get mentioned. The publication of failures can potentially shorten the learning curve of research because they are discoveries in their own right. Here we share those things that have not worked in our soft sensing deep learning experiments in the hope that future researchers and application engineers can learn from these lessons. This article is compiled from our experiments with multiple deep learning algorithms including Autoencoder, CNN, GNN and Transformer to predict the quality control variables in the wafer manufacturing process in Seagate, based on the sensor data collected during deposition and etching processes. We experimented on various pre-processing and modeling methods and here are some of the methods that did not show a good performance: using focal loss or hinge loss to replace cross entropy, using standard or robust scaler to replace min-max scaler, scaling the data by groups, using oversampling methods such as random sampling or SMOTE, and selecting features based on domain knowledge. We hope that these experiments will give some intuition to future researchers and save their time for experiments. We recognize that these conclusions are drawn from our experiments only, and accept that it is possible that these things may work in a different context. We welcome all kinds of discussion or collaboration.