A tandem neural network has been developed that can infer key physical parameters of semiconductor materials from simple transistor measurements, researchers at the Tokyo Institute of Science report. While traditional approaches to this type of analysis take hours or even days, the proposed system produces results in less than 1 millisecond and with near-perfect accuracy.
Inverse estimation of semiconductor properties using tandem neural networks
Modern electronics relies heavily on semiconductor devices, whose performance is determined by material properties such as defect density and charge transport properties. Although it is now relatively easy for engineers to measure the behavior of transistors, it is still much more difficult to determine the fundamental material properties that cause that behavior. This type of “inverse” analysis is essential to developing better electronics and improving manufacturing processes, making it increasingly important to find efficient ways to probe semiconductor materials.
One of the main difficulties when dealing with such inverse problems is related to what scientists call “multivaluedness.” Simply put, different combinations of material properties result in nearly identical transistor properties, making it extremely difficult to work backwards, i.e., to determine the physical properties of semiconductor materials based solely on device performance measurements. Traditional approaches typically rely on running computer simulations and trial-and-error optimization, which can take hours or even days. However, what would happen if we used machine learning to tackle multi-value processing?
In a recent study, a research team led by then graduate student Masatoshi Kimura, assistant professor Keisuke Ide, and professor Toshio Kamiya from the MDX Element Strategy Research Center at the Tokyo University of Science, collaborated with Yokohama City University in Japan and National Sun Yat-sen University in Taiwan to address this issue. As reported in a paper published in an online journal advanced intelligent system May 27, 2026 They developed a new machine learning framework that can solve inverse problems incredibly fast.
The team’s approach focuses on what is known as a tandem neural network (TNN), which is essentially two machine learning models linked in series. The first model attempts to solve the inverse problem by estimating material properties from transistor measurements. The second model is a pre-trained forward network that uses the material estimates produced by the first model to reconstruct the original transistor characteristics. By using the output of this second model as part of the training input given to the first model, the entire system learns to find solutions that are mathematically reasonable and physically consistent.
The researchers trained this TNN using a dataset of 1,000 amorphous indium gallium zinc oxide (a-IGZO) transistors covering six important physical parameters, including defect density, trap state properties, and electron mobility. Even though the parameter range was about 1,000 times wider than in previous machine learning studies, the TNN was able to infer all six parameters from a single current-voltage curve in less than 1 millisecond with near-perfect accuracy.
The team also tested the system under five different conditions using real transistors manufactured in the lab. The model was able to successfully reproduce the measured behavior without any additional optimization steps or adjustments. “Compared to traditional device simulation-based methods, which require hundreds of iterations and take tens of hours to days, the proposed approach achieved speedups of more than six orders of magnitude,” said Ide.
The features of the proposed architecture demonstrate some concrete applications. For example, in manufacturing, this approach can be used to perform instant quality checks on transistors coming off the production line. Meanwhile, in research settings, AI agents could serve as core tools in autonomous laboratory systems to design, run, and analyze experiments with minimal human input. There may also be use cases in other fields such as materials science, chemistry, and optics, as Ide concludes, “We expect that our approach can be applied not only to semiconductors, but also to a variety of inverse problems involving multivalued properties.”
- author:
- Masatoshi Kimura1Keisuke Ide1*, Zhou Kuanju2Jun Shimizu1Takayoshi Katase1,3Hidenori Hiramatsu1,3Kei Terayama1,4Hideo Hosono1Toshio Kamiya1*
*Responsible author - title:
- Tandem neural networks quickly solve multivalued inverse problems: Application to characterization of oxide semiconductors
- journal:
- advanced intelligent system
- Affiliation:
- 1Tokyo University of Science Research Institute MDX Element Strategy Research Center
2Department of Physics, National Sun Yat-sen University (Taiwan)
3Tokyo University of Science Research Institute Materials and Structure Research
4Yokohama City University Graduate School of Life and Medical Sciences
