Search for a command to run...
Wire bonding quality is critical to semiconductor reliability. Introducing machine learning for real-time prediction, early warning systems, and comprehensive defect coverage becomes particularly essential with advanced manufacturing trends. However, existing machine learning approaches suffer from data limitations of amount and diversity, noise interference, and an inability to capture physical dynamics, resulting in a bottleneck to breakthrough the prediction performance. In this paper, we propose a novel Physics-Informed Ensemble (PIE) Learning framework that derives physical relationships through designed experiments, employs physics-driven data augmentation to address data limitations, and leverages adaptive ensemble strategies to dynamically balance model strengths across diverse operational scenarios. Our experiments focus on ball shape parameter performance: Ball Thickness, Ball Size X and Ball Size Y. The framework maintains average prediction deviations within <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.92 \mu \mathrm{m}$</tex>, with the prediction deviation of Ball Thickness kept within <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.61 \mu \mathrm{m}$</tex> for in range variations, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.32 \mu \mathrm{m}$</tex> for out-range variations. Compared to a single model approach, PIE delivers an improvement in prediction accuracy of over 14%, achieving an error reduction of up to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.8 \mu \mathrm{m}$</tex>. Our framework enables high-precision wire bonding quality prediction while enhancing robustness to process variations and unforeseen operating conditions, thereby improving the reliability of semiconductor packaging quality control systems.