Abstract:The Redistribution Layer (RDL) is an essential component in modern integrated circuits (ICs) which mainly conducts electrical signals from the I/O to the internal circuits on chips. The cause of RDLs’ fracture is high stress during packaging. To estimate the stresses in packaging, Finite Element Analysis (FEA) is frequently used. However, the structure of RDL, especially the complex Cu lines, is hard to model. The mostly-used effective approach is to simplify the layer to a homogeneous one by volume percentage method. This approach greatly reduces the complexity of modeling the internal geometry. However, the volume average algorithm ignores the endowed directionality by converting the layer to be isotropic, meaning that it is constrained to simple symmetric patterns. Since Cu wiring patterns cannot guarantee directional uniformity in massive production, layers have differences in orthogonal mechanical performance. Although simulation-based methods can accurately obtain the orthotropic properties, it is relatively time-costly and the cost will increase significantly with the decrease of line width and the increase of wiring quantity.This paper proposes convolutional neural networks (CNNs) for calculating the orthotropic material properties of the equivalent layer based on effective modeling. Material properties, including Young’s modulus, tangent modulus, yield stress, Poisson’s ratio, shear modulus, and CTE along each principal axis can be obtained by analyzing the wiring pattern using an asymmetrically-kernelized CNN. The major steps are constructed as follows. Firstly, the properties concerned are obtained through simulation by effective modeling. Patterned models are built and the simulation of uniaxial tensile and free thermal expansion tests are carried out by FEM. The required parameters can be afterward obtained in the virtual testing. Secondly, the obtained data are divided into training sets and validation sets at a ratio of 9 to 1. Finally, CNNs are established for training. The Mean Absolute Percentage Error (MAPE) is adopted as the custom loss. CNN training results show that the network can greatly reduce the time and RAM cost required for simulation while ensuring good prediction accuracy (MAPE<10%).Based on the method, the warpage of multi-layer was investigated. A 5-layer model exhibits asymmetric warpage during the reflow process, caused by the RDL’s orthotropy. Compared to the refined model of the multi-layer structure using 3D solid elements, the present method only needs to designate the mechanical properties (output of trained CNN with patterns as input) of each layer to 2D simplified shell elements, greatly reducing computation time. The deviation between the predicted warpage by CNN and the FEM simulation of the 5-layer wafer is within 10%.

The Prediction of Orthotropic Material Properties for RDL Based on Effective Modeling and CNN

Machine Learning and Materials Informatics Approaches for Predicting Transverse Mechanical Properties of Unidirectional CFRP Composites with Microvoids

An RDL Modeling and Thermo-Mechanical Simulation Method of 2.5D/3D Advanced Package Considering the Layout Impact Based on Machine Learning

A Multi-channel Input Framework for Structure Displacement Response Prediction Using Convolutional Neural Network

Bridging finite element and deep learning: High-resolution stress distribution prediction in structural components

Employing Deep Learning in Non‐parametric Inverse Visualization of Elastic–plastic Mechanisms in Dual‐phase Steels

Data-Driven Stress/Warpage Analyses Based on Stoney Equation for Packaging Applications

Tensile Performance Prediction of CFRPs with Voids Using Multiscale Analysis and Neural Networks

Efficient mapping between void shapes and stress fields using Deep Convolutional Neural Networks with Sparse Data

Comparison of Mechanical Modeling to Warpage Estimation of RDL-First Fan-Out Panel-Level Packaging

CmpCNN: CMP Modeling with Transfer Learning CNN Architecture.

Prediction of local elasto-plastic stress and strain fields in a two-phase composite microstructure using a deep convolutional neural network

A deep learning convolutional neural network and multi-layer perceptron hybrid fusion model for predicting the mechanical properties of carbon fiber

Tool Wear Prediction in Milling CFRP with Different Fiber Orientations Based on Multi-Channel 1DCNN-LSTM

Rapid and accurate predictions of perfect and defective material properties in atomistic simulation using the power of 3D CNN-based trained artificial neural networks

Prediction of mechanical solutions for a laminated LCEs system fusing an analytical model and neural networks

Material Performance Prediction and Mechanical Structure Optimization Design Algorithm Based on Multi-Source Heterogeneous Data Fusion

Prediction of microstructural-dependent mechanical properties, progressive damage, and stress distribution from X-ray computed tomography scans using a deep learning workflow

Warpage Simulation For Multilayered Mis Substrate By Application Of Image Recognition Method

Failure Life Prediction of Wafer Level Packaging using DoS with AI Technology

Three-dimensional convolutional neural network (3D-CNN) for heterogeneous material homogenization