Reliability of high-power LEDs remains an issue in the automotive industry due to the high thermo-mechanical mismatch between ceramic die carrier and the metal core PCBs used for heat management: Package and solder pad design strongly impacts this. However, a lack of data causes issues comparing diverse package designs. Building on a dataset of 1400 LEDs across seven ceramic packages, this study leverages convolutional neural networks (CNNs) to predict the lifetimes of different LEDs. It integrates FEA data to predict the lifetime. Current state of the art is based on the Coffin Manson and Syed Model, which are based on Power Law and used selective FEA data points [1][2]. 

Thermo-mechanical simulations of temperature shock cycles from -40°C to 125°C were undertaken on seven ceramic LEDs. The research employs the Garofalo creep model for SAC305 and SAC105, with the simulation processed into metrics for stress, creep strain, and dissipated creep energy in the solder layer. Incorporating Transient Thermal Analysis (TTA) data to access the degradation of the solder interconnect from 440 samples, cycles to failure were determined. The choice of these 440 samples was guided by the availability of the Garofalo model, which is specific to two solders: SAC105 and SAC305. 2D grid-based interpolation and averaging addresses the complexity and voluminous nature of 3D Ansys data, streamlining it into a more manageable 2D format. By emphasizing uniform resolution and employing linear interpolation, this method offers enhanced accuracy in capturing the detailed stress, strain, and energy values compared to state of art which selected individual values, thereby refining the overall fidelity of analytical outcome.

Inputs were initial Scanning Acoustic Microscopy, X-ray, and FEA grid data, with the output describing Cycles to Failure for the sample. A simple Dense neural network yielded an R2 score of 0.20. However, integrating CNNs with FEA, SAM, and X-ray data augmented accuracy. By employing L2 regularization and Dropout techniques, an R2 score of 0.65 was realized as shown in Figure 1, highlighting potential enhancements with alternate solder creep predictive models in the future. This integration of 2D FEA grids with CNNs provide a pioneering method for solder joint lifespan predictions, charting avenues for further exploration.