Professor Jeong Chang-uk's team at the Ulsan National Institute of Science and Technology (UNIST) Graduate School of Semiconductor Materials and Components said on the 26th it developed a new artificial intelligence (AI)-based data correction algorithm that can predict the paths where heat spreads and the locations where stress concentrates during semiconductor processing or packaging.
The research was accepted to the International Conference on Learning Representations (ICLR) 2026, regarded as one of the three major international conferences in artificial intelligence. Heat, known as the enemy of semiconductors, can lead to performance degradation, cracks, or damage if it concentrates in specific areas during processing or is difficult to control.
The researchers developed a "pi (π)-invariant test-time correction" algorithm that realigns new input data to match the criteria of previously learned data, and applied it to semiconductor processes. The algorithm converts "out-of-distribution inputs" into familiar in-distribution forms while obeying physical laws.
When new input data arrive, the method first uses the π value to find the physically most similar data among the previously learned data, adjusts the conditions to be similar, and only then feeds them into the AI model for computation.
In particular, it is cost-effective because it can be attached to and used with existing AI models without additional retraining, and the team said they reduced computational load by grouping similar data and comparing only representative values instead of comparing all training data one by one.
The team applied the algorithm to heat conduction and linear elasticity problems. Tests showed stable predictions even under new conditions that existing models struggled with. The mean absolute error decreased by up to about 91%.
A member of the research team said it will be used to reduce computation time and expense in various engineering simulations where scale and conditions keep changing, including thermal design of semiconductor chips, package reliability assessment, battery thermal management, and structural safety analysis.