A Methodology to Optimize Laser Dicing Parameters to Maximize Dicing Quality Through Machine Learning

Abstract

Laser dicing has been shown to improve the die strength for thin dies compared to standard machine dicing processes. But like blade dicing, or any separation process, the laser dicing process needs to be optimized for a given substrate to obtain the best dicing quality. Optimization of laser dicing process parameters is challenging because the dicing quality can be influenced by multiple parameters, which can either be independent or co-dependent. Practically, in the absence of structured studies, each laser parameter needs to be varied and the output dicing quality is required to be tested to obtain an acceptable dicing process window for any substrate. Such a method needs to be repeated to create new recipes for any change in die thicknesses, back-end-of-line (BEOL) configurations, etc. This approach is ineffective, sub-optimal and time-consuming. In this study, a methodology is developed to systematically study the effect of key laser dicing process parameters on dicing quality using a machine learning algorithm. First, a design of experiments (DoE) is modeled to study six input parameters with ten levels. The levels are carefully designed based on experience, past data, and tool limitations in an effort to further push the boundaries of the current process window. The design space for training and test data is generated using space filling algorithm such as Latin Hypercube. The dicing quality measures such as dicing width and die strength are measured for each iteration and treated as output. Each experiment is repeated 15 times and the mean value of the output parameter is used for training a machine learning model. A random forest-based machine learning algorithm is used to create a surrogate model relating the input and output parameters. Such a model is then used to understand the interactions between input parameters and identify the optimal process window maximizing dicing quality for various sets of input parameter settings. The optimized process parameters are then validated experimentally. The optimized process parameters resulted in increasing the dicing quality compared to the baseline appreciably. This methodology can be extended to any complex, multi-factor dependent separation process to find the optimized process window.