Elucidating microbubble structure behavior with a Shapley Additive Explanations neural network algorithm

Silica microresonators (microbubbles) are considered excellent candidates due to the realization of ultrahigh quality factors in whispering gallery mode resonators (WGMs), which can confine significant optical powers in small spaces. The challenge in the optimal design of microbubbles is to calculate their unique properties and enhance their capabilities as devices by understanding their physical mechanisms. Machine learning (ML) strategies have been employed for microbubble design. However, these approaches are often considered ‘black boxes’ due to the model’s lack of explanations for their predictions. This study introduces a feedforward neural network (FFNN) model that accurately forecasts the optical properties of microbubbles. Utilizing the SHAP (Shapley Additive Explanations) method, an analytical tool offering explanations, we delineate the precise impact of microbubble geometric parameters on the predictions of FFNN model and pinpoint the critical factors influencing their optical properties. By employing reverse engineering, we can deduce the geometric parameters of microbubbles from desired outcomes, thus providing an approach to the optimal design of these structures. This research not only equips us with a powerful instrument for a nuanced comprehension of microbubble structures and performance optimization but also paves new avenues for exploration in the realms of optics and photonics. •FFNN models accurately predicted microbubble optical properties, achieving R2 up to 0.996.•SHAP algorithm analyzed the impact of microbubble geometry on optical properties.•SHAP insights revealed complex relationships between optical properties and geometry.•The method achieved a 0.01865 error by comparing predicted features with COMSOL results.•This approach provides a systematic framework for optimizing microbubble structures.