Deep-Learning for Semiconductor Process Development Optimization

Reducing time taken for Design-of-Experiments (DoEs) through Deep-Learning

Complex processes involved in synthesis of semiconductor materials are typically cumbersome, and involves optimal setting of several interdependent parameters to achieve the desired material characteristics. A number of parameters such as the applied voltage and its time period, and other environmental factors significantly influence the characteristics of a fabricated semiconductor device. Understanding the electronic characteristics of a newly fabricated device, and estimating the optimal setting of the several process parameters to achieve a desired characteristics, are therefore, very challenging.
 
Researchers in semiconductor fabrication labs would require to conduct several traditional Design-of-Experiments (DoEs) over a long time period, ranging from several days to several months, to estimate this optimal setting for each synthesized material.  

In A*STAR Institute for Material Research and Engineering (IMRE), experiments of such will be carried out repeatedly on each fabricated photo-CELIV device under varying conditions such as Laser pulse intensity, Pulse width, Offset, Delays, Time of exposure to air, etc. 

To assist them in the reduction of this tedious repetitive process and time taken for experiments, A*STAR Institute for Infocomm Research (I2R) embarked on the development of advanced regression models to replace the experimentation of different conditions and parameters.

With A*STAR IMRE’s historical data of the devices and experiments carried out since the beginning of production, A*STAR I2R is able to train these advanced regression models to predict the conductivity of a device and the optimal parameters for a desired conductivity, based on a few experimental conditions. 

Experiment under conditions – Time of exposure to air & Varying Voltages

Experiment 1: Characterize the device under varying air exposures
In this experiment, we characterize the device at varying periods of air exposures. For example, we train the deep learning model with data obtained from the device exposed to air for 1 hour, and predict the characteristics of the device under 30 min air exposure. The model predicts the conductivity characteristics of the device at varying air exposure condition, with error of the order of 10-5.

 

Experiment 2: Characterize the device under varying applied voltage
This experiment is aimed at defining the characteristics of the device at varied levels of applied voltage. For example, we train a deep regression model with data obtained by subjecting the device to voltage levels of 1V, 1.5V, 2V and 2.5V, each at different periods of air exposure. We then predict the characteristics of the device at an applied voltage of 3V, under air exposure of 30min and 1hour. The model predicts the conductivity characteristics of the device at varying voltage levels with an error of the order of 10-7.

Thus, it can be observed that through such accuracy of the models, projections could be made on a higher range of parameters (eg: higher voltages) without subjecting the actual device to their physical limitations and accidental damages.

These models have so far garnered 98.4% accuracy in prediction. This could translate to a reduction of at least five (5) times of the original DoEs (Design of Experiments) time taken for the researchers. While utilising these advanced regression models, the team also discovered the additional characteristics of the materials which were not highlighted through the conventional experimentation methods.