AI-Based Power Prediction Control of Magnetron Sputtering High-Voltage Supply

 

The discharge in a magnetron sputtering chamber is a complex, time-varying process. The plasma impedance changes with the gas pressure, the target erosion, the magnetic field, and the temperature of the chamber. The conventional feedback loop reacts to these changes after the changes occur, and the response time of the loop limits the speed with which the power can be corrected. The AI-based prediction control uses a model of the process to forecast the power requirement ahead of time, so the correction can be applied before the disturbance affects the film.

 

The prediction model is trained with the data collected during the process operation. The data include the gas flow, the pressure, the target voltage and current, the magnetic field setting, and the resulting film properties. The model learns the correlations between these parameters and the discharge power, and the model is updated continuously with the new data. The training and the update are performed by the control system, which runs the model in parallel with the conventional feedback loop.

 

The prediction control provides the feed-forward component of the regulation. The model receives the setpoint changes and the measured process parameters, and the model predicts the required output power for the next interval. The predicted value is combined with the feedback error, and the combined command drives the power stage. The feed-forward component reduces the response time of the power regulation, and the feedback component corrects the residual error and the model inaccuracy.

 

One of the main sources of power variation in magnetron sputtering is the target erosion. As the target is consumed, the discharge impedance changes and the power drifts. The prediction model learns the erosion trend from the accumulated operating time and the measured parameters, and the model adjusts the power command to compensate for the drift. The compensation maintains the deposition rate constant over the life of the target, improving the uniformity of the film across the batches.

 

Another source of variation is the arcing of the discharge. The arc events cause a rapid drop of the power, and the film quality is affected if the power is not restored quickly. The prediction control detects the onset of the arc from the fast changes of the measured parameters and prepares the power stage for the recovery. The recovery sequence is optimized by the model, and the time during which the power is reduced is minimized.

 

The AI-based control also supports the optimization of the process parameters. The model can be used to explore the relationship between the process settings and the film properties, and the control system can suggest the settings that give the target film quality with the minimum power consumption. The optimization is performed offline with the historical data, and the recommended settings are validated by the test runs before application to the production.

 

The prediction control is implemented on the control platform of the supply, which runs the model inference in real time. The platform is designed with the computing resources that support the model without affecting the deterministic behavior of the control loops. The model is versioned, and the updates are validated before deployment to the production systems. The validation includes the comparison of the predicted power with the measured power over a test period.

 

The data management is an important part of the AI-based approach. The process data are stored in a central database that is shared among the coating lines, and the model is trained with the data from all the lines. The model therefore learns the common behavior of the process and the specific characteristics of each line. The model is retrained at defined intervals, and the performance of the model is monitored continuously.

 

The film quality is verified by the metrology of the coated samples. The thickness, the optical properties, and the adhesion of the films are measured, and the measurements are correlated with the power records of the coating runs. The correlation validates the benefit of the prediction control, and the improvement of the film uniformity is quantified. The quantified benefit justifies the investment in the AI-based control.

 

The implementation of the prediction control follows a staged approach that reduces the risk of the introduction. In the first stage, the model runs in parallel with the conventional control and the predicted values are recorded without being applied to the power stage. The comparison of the predicted values with the measured values validates the accuracy of the model over the full range of the operating conditions. In the second stage, the feed-forward path is enabled for a single coating line, and the film uniformity of the line is compared with the uniformity of the lines that use the conventional control. In the third stage, the validated control is deployed to the remaining lines. The staged approach provides the evidence base for each deployment decision and preserves the production stability during the transition. The experience gained in each stage is fed back into the model training, and the continuous improvement of the model is supported by the growing data set.

 

In summary, the AI-based power prediction control of the magnetron sputtering high-voltage supply integrates a learning model with the conventional feedback regulation, providing a feed-forward compensation that anticipates the process variations. The result is a supply that maintains the deposition power required for uniform films while adapting to the changes of the target, the gas, and the chamber conditions. Every improvement of the prediction model, every refinement of the feed-forward path, and every enhancement of the data management contributes directly to the quality and the efficiency of the coating process. The engineering effort continues as the sputtering processes become more complex and the film specifications become more demanding.