Abstract
Electronic packaging has become a critical technology due to market demand. Fan-Out Panel Level Packaging (FO-PLP) has become the industry's most popular packaging due to its large area and good use ratio. It is important to consider warpage issues during molding. Excessive warpage will make the next process impossible. Measuring the warpage of panel-level packages with different geometries has become one of the important tasks. Testing the warpage in an experiment way will take a lot of time. Many researchers use simulations instead of experiments to save time and cost. In this study, we use the Finite Element Method (FEM) instead of experiments. When simulating the Fan-out panel level package, we build a model for a variety of geometric shapes, and then we apply the boundary conditions and thermal loading. The equivalent CTE for the molding compound, which simplifies the complex material properties of the compound. In the end, we will measure the Z-displacement as the warpage value. Although the simulation takes less time than the experiment, FEM still requires a lot of calculation time and verification. With the improvement of computing equipment, artificial intelligence has become a popular method to estimate value and avoid human error. We use FEM to build warping datasets of different geometries for machine learning. Machine learning builds a trained model with the training data, and the trained model can estimate the value immediately. For machine learning, we choose Modified Convolution Neural Network as the learning algorithm. In Modified Convolution Neural Network, the filter will detect the edge of the warpage model. [1] The algorithm can select the important data points as the training data. Edge detection will reduce machine learning's overall calculation amount and ensure enhanced accuracy. An application of Particle Swarm Optimization [2] results in an improvement in learning performance and avoidance of local minima. The algorithm will use the wisdom of the group to help find the initial weight and bias to improve the accuracy of the estimation.