Informatics
Semiconductor Internal Analysis That Used to Take Days Now Takes Under One Millisecond with AI
Key points (reviewing AI explanation)
By combining two neural networks, researchers shortened semiconductor internal analysis that used to take several days to under one millisecond. The training data was about 400 cases, roughly one-hundredth of the conventional amount.
Explanation
Analysis that works backward from semiconductor measurement results to infer the internal state has sometimes taken several days. Assistant Professor Keisuke Ide of Institute of Science Tokyo and colleagues succeeded in performing this analysis in under one millisecond by combining two neural networks (announced August 25, 2026).
The mechanism has two stages. The first network estimates the internal state of the material from the measurement results, and the second network checks whether the original measurement results can be reproduced from that estimated state. Rather than simply outputting an answer, the system is characterized by verifying the answer in reverse.
The training data consisted of about 400 cases, roughly one-hundredth the amount used in previous research, yet still achieved high accuracy. Beyond rapid analysis of measurement data at semiconductor plants, applications are expected in materials science, chemistry, optics, and other fields involving "working backward from results to causes."
Source: www.isct.ac.jp