Towards Reliable Parameter Extraction in MEMS Final Module Testing Using Bayesian Inference.

Area of Science:

• Engineering
• Computer Science
• Data Science

Background:

• Micro-Electro-Mechanical Systems (MEMS) testing demands high precision and reliability.
• Machine learning (ML) offers runtime efficiency but faces challenges in uncertainty quantification and reliability guarantees.
• Existing ML methods require robust uncertainty estimation for trustworthy MEMS testing.

Purpose of the Study:

• To present a novel machine learning approach for MEMS testing using Bayesian inference.
• To evaluate the trustworthiness of ML estimations through uncertainty quantification.
• To enhance the reliability and efficiency of parameter estimation in MEMS module testing.

Main Methods:

• Bayesian neural network (BNN)
• Mixture density network (MDN)
• Probabilistic Bayesian neural network (PBNN)
• BayesFlow
• Evaluation under varying training set sizes, noise levels, and out-of-distribution conditions (damping factor variation).

Main Results:

• BayesFlow consistently outperformed other methods in predictive performance.
• The study evaluated both epistemic and aleatoric uncertainties.
• PBNN enabled the distinction and analysis of different uncertainty types.

Conclusions:

• The proposed Bayesian inference approach enhances trustworthiness in ML-based MEMS testing.
• BayesFlow shows significant promise for reliable and efficient MEMS parameter estimation.
• Thorough uncertainty evaluation is crucial before deploying ML models in MEMS testing.