DEVELOPMENT OF A VIRTUAL VACUUM CHAMBER PROTOTYPE USING MACHINE LEARNING

Abstract

This work aimed to present the modeling, development, and validation process of a vacuum chamber virtual prototype capable of incorporating experimental uncertainties characteristic of real-world testing, with the purpose of accurately predicting the time required for the pumping process using machine learning. To train and validate the Random Forest model, a synthetic dataset containing 125,000 combinations of these variables was generated, with Gaussian noise added to the pumping time to simulate instrumental uncertainties. Performance was evaluated using mean absolute error and the coefficient of determination (R2), in addition to uncertainty estimation for the predictions and parametric optimization to achieve physically feasible target times. The results indicated high precision (R2 close to 1 and mean absolute error below 0.05 s), strong generalization capability, and fidelity in noise representation. The simulator proved promising to support the design of experiments. As limitations, it is acknowledged that noise was applied only to the model's output, without considering uncertainties in the input variables, and that the underlying physical model disregards phenomena such as leaks or non-ideal flow regimes, which may restrict its applicability in more complex experimental scenarios. It is worth noting that this work constitutes a methodological proof-of-concept study whose validation was performed in a controlled environment using synthetic data, establishing the foundation for future extensions involving experimental validation and the incorporation of more complex scenarios.