A tandem neural network capable of inferring key physical parameters of semiconductor materials from simple transistor measurements has been developed, as reported by researchers from Institute of Science Tokyo. While conventional approaches for this type of analysis require hours or even days, the proposed system produces results in under one millisecond with near-perfect accuracy.
Modern electronics relies heavily on semiconductor devices, whose performance is shaped by material properties like defect density and charge transport characteristics. Even though it’s relatively easy today for engineers to measure how a transistor behaves, determining the underlying material properties responsible for that behavior remains far more difficult. This type of “inverse” analysis is essential for developing better electronics and improving manufacturing processes, and thus finding efficient ways of probing semiconductor materials is becoming increasingly important.
One of the main difficulties when dealing with such inverse problems is related to what scientists call “multivaluedness.” Simply put, because different combinations of material properties can produce nearly identical transistor characteristics, it is quite challenging to work backwards—that is, identify the physical properties of a semiconductor material based solely on device performance measurements. Conventional approaches typically rely on running computer simulations and trial-and-error optimization, which can take hours or even days. But what if we leveraged machine learning to tackle multivaluedness?
In a recent study, a research team led by graduate student (at the time) Masatoshi Kimura, Assistant Professor Keisuke Ide, and Professor Toshio Kamiya from the MDX Research Center for Element Strategy, Institute of Science Tokyo, Japan, in collaboration with Yokohama City University, Japan, and National Sun Yat-sen University, Taiwan, tackled this challenge. As reported in their paper, which was published online in the journal Advanced Intelligent Systems on May 27, 2026, they developed a new machine learning framework that can solve inverse problems incredibly fast.
The team’s approach centers on what’s known as a tandem neural network (TNN), which is essentially two machine learning models linked in series. The first model tries to solve the inverse problem, 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 overall system learns to find solutions that are both mathematically plausible and physically consistent.
The researchers trained this TNN using 1,000 amorphous indium–gallium–zinc oxide (a-IGZO) transistor datasets, covering six important physical parameters like defect densities, trap-state characteristics, and electron mobility. Even though the parameter range was roughly 1,000 times wider than previous machine learning studies, the TNN could infer all six parameters from a single current–voltage curve in under one millisecond with near-perfect accuracy.
The team also tested the system using real transistors fabricated in the laboratory under five different conditions; the model could successfully reproduce their measured behavior without any additional optimization steps or adjustments. “Compared with conventional device-simulation-based methods requiring hundreds of iterative calculations and taking tens of hours to several days, the proposed approach achieved a speedup of over six orders of magnitude,” remarks Ide.
The capabilities of the proposed architecture point to several concrete applications. For example, in manufacturing, this approach could be used to perform instant quality checks on transistors as they come off a production line. Meanwhile, in research settings, it could serve as a core tool for autonomous laboratory systems where AI agents design, run, and analyze experiments with minimal human input. There could even be use cases in other fields like materials science, chemistry, and optics, as Ide concludes: “We expect our approach to be applicable not only to semiconductors but also to a wide variety of inverse problems involving multivaluedness.”
***
About Institute of Science Tokyo (Science Tokyo)
Institute of Science Tokyo (Science Tokyo) was established on October 1, 2024, following the merger between Tokyo Medical and Dental University (TMDU) and Tokyo Institute of Technology (Tokyo Tech), with the mission of “Advancing science and human wellbeing to create value for and with society.”
Advanced Intelligent Systems
Experimental study
Not applicable
Tandem Neural Network Rapidly Solves Multivalued Inverse Problems: Application to Oxide-Semiconductor Characterization
27-May-2026
The authors declare no conflicts of interest.