Recursive Transformers Boost Semiconductor Reliability Modeling

A new study investigates the application of recursive transformer paradigms as surrogate models for complex engineering simulations, specifically focusing on the thermo-mechanical reliability of advanced semiconductor packages. Traditional transformer architectures often suffer from over-parameterization when applied to the small, low-dimensional datasets typical in engineering design, leading to overfitting and unnecessary computational overhead. The research systematically compared three recursive transformer types: Tiny Recursive Model, Depth Recursive transformer, and a simple recursive transformer. The evaluation focused on their predictive performance, parameter count, and computational complexity, providing practical guidelines for selecting architectures under resource constraints. These recursive weight-sharing transformers demonstrated an effective balance between prediction accuracy, parameter efficiency, and computational cost. This makes them particularly suitable for scenarios where generating large simulation data is expensive, such as evaluating stress and warpage from thermal cycling in semiconductor design, or solving Laplace PDE for capacitance fields.