The modern world relies heavily on semiconductor manufacturing, and the industry is highly competitive and complex. Semiconductor foundries have to perform numerous intricate steps that demand high precision and quality control to maintain their Manufacturing Excellence.

Fabs also contend with rapidly increasing demands for high-performance chips and intensifying external stressors on operational performance. The pressure for semiconductor manufacturers to optimize their production processes is palpable. 

A proactive approach to process optimization leverages advanced technologies such as artificial intelligence (AI) and machine learning (ML). Semiconductor manufacturers who apply them intelligently can mitigate the impact of adverse externalities. Data-driven optimization allows fabs to remain competitive in an increasingly dynamic and unpredictable marketplace. 

Below is the optimization journey of a leading manufacturer of semiconductors, LEDs, and other advanced materials. Read on to explore:

  1. how this fab introduced data-driven solutions to improve predictions.
  2. the measurable value of digital-era process optimization.

Understanding The Challenge 

In the targeted process, the semiconductor manufacturer used ion beams to etch silicon wafers for specific thickness profiles. The procedure involved directing the beam through a grid to disperse it over the wafer and create the desired etch profile. However, even if the grid measurements were within specifications, the resulting etch profile left room for quality improvement. 

Specifically, the existing baseline model predicted wafer uniformity but not the two-dimensional etch profile. Therefore, grids could be deemed acceptable despite being defective. Factored in with the complexity and non-linearity of dispersion, our data scientists identified an optimization opportunity. 

Addressing this yield detractor, DataProphet applied a machine-learning model. It would determine the quality of the resulting wafers based on grid measurements.

In other words, the contextualized challenge was to develop a more effective baseline model architecture. The goal was to improve the training pipeline for more accurate prediction of the following metrics:

  • Wafer uniformity.
  • The wafer’s two-dimensional etch profile. 

Optimizing Predictions with AI and Data Expertise 

Our data scientists led the project in consultation with in-house subject matter experts. Its success would demonstrate improved predictions in the semiconductor manufacturer’s silicon wafer etching process.  

A new architecture to address the baseline model’s limitations was proposed. Our data scientists fed input parameters into a supervised deep-learning model during the multiple etching steps. Convolutional and recurrent layers captured the spatial distribution of input points and the temporal nature of the etching process. 

Data augmentation techniques such as rotation, scaling, and flipping of input channels were also used. They improved the training pipeline to increase the size and diversity of the training set. Further, the team implemented early stopping and dropout regularization. This technique prevented overfitting and improved generalization.

Testing the Results of the Optimizing Predictions 

In practice, the model predicted the wafer’s two-dimensional etch profile in addition to uniformity, paving the way for more accurate quality control. Evaluated on a large dataset, the new model significantly outperformed the baseline model’s predictive capability. Both wafer uniformity and etch profile prediction improved.

However, to clarify the outcome, a different technique was necessary. To this end, our data science team performed an ablation study. It evaluated the relative contributions of the new model to the overall performance. The ablation study proved that convolutional and recurrent layers were essential for capturing the spatial and temporal patterns in the process data.

Our machine learning specialists had achieved something unprecedented. As DataProphet’s CTO (and this project’s spearhead) Benny Leonard puts it:

“We predicted deformities in the shape of the chip post etching within the order of six angstroms — this metric constituted extreme accuracy compared to the previous benchmark.” 

DataProphet’s etch quality prediction solution gave the semiconductor manufacturer valuable insights for optimizing its silicon wafer etching process.

illustration of DataProphet's AI solution's impact on semiconductor manufacturing process optimization

Future Implications for Semiconductor Manufacturing Process Optimization

The success of this project highlights the importance of investing in data expertise and adopting a proactive approach toward process optimization in the semiconductor manufacturing industry.

Approaching process optimization with contextualized data-driven solutions is a value-enabling tool the semiconductor manufacturing industry can leverage across multiple process steps. As shown, utilizing advanced technologies like AI and ML to improve predictions helps chipmakers make sense of their data. 

By embracing AI and machine learning at the next level, semiconductor manufacturers can unlock new insights into their production processes to make data-driven decisions that optimize quality, reduce costs, and improve performance. At scale, data augmentation techniques geared for process optimization also embed much-needed momentum for fabs entering the digital era of manufacturing.