In a breakthrough that could revolutionize semiconductor analysis, a team of researchers at the Tokyo University of Science has pioneered a tandem neural network (TNN) architecture that dramatically accelerates the inference of critical material parameters from transistor measurements. This new machine learning framework addresses one of the most vexing challenges in semiconductor physics, the multivalued inverse problem, achieving unprecedented speed and accuracy over traditional computational methods.
Semiconductor devices power nearly every aspect of modern electronics, from smartphones to advanced computing systems. The performance and reliability of these devices depend on complex material properties such as defect density, trap states, and charge carrier mobility. Although engineers can quickly measure transistor behavior through current-voltage characteristics, deciphering the precise material properties that cause this behavior remains an elusive inverse problem. Traditional approaches rely heavily on iterative simulation and optimization techniques, which often take hours or even days to converge on a reasonable solution.
At the heart of the difficulty lies the phenomenon of polyvalence. That is, different combinations of semiconductor material parameters can result in virtually indistinguishable transistor responses. This ambiguity complicates the task of tracing device performance back to its physical origins. The research team recognized that traditional simulation frameworks do not have the ability to efficiently avoid this degeneracy. Their insight was to harness the power of tandem neural networks, an innovative deep learning configuration that can simultaneously enhance physical consistency and mathematical validity in parameter estimation.
The tandem neural network developed by the team combines two interconnected models operating in series. The first inverse model proposes a candidate set of material parameters based on the observed transistor behavior. This output is input into a pre-trained forward model to reconstruct transistor characteristics from material parameters. By integrating the forward model reconstruction into the training loss function of the inverse network, the system effectively teaches itself to generate solutions that not only fit the measured data, but also fit the physical laws governing transistor operation.
Training this sophisticated architecture required an extensive dataset of more than 1,000 simulated transistor measurements, with a particular focus on amorphous indium gallium zinc oxide (a-IGZO) transistors, a material of great interest for flexible and transparent electronics. This dataset spanned a parameter space approximately 1,000 times wider than previous studies, encompassing six key variables: defect density, electron trapping states, and carrier mobility between them. Remarkably, the TNN estimated all six parameters from a single transistor’s current-voltage curve in less than 1 millisecond, achieving near-perfect accuracy over the entire range.
Beyond simulation, real-world validation was paramount. The researchers fabricated a series of a-IGZO transistors under five different process conditions in the lab. When performing TNN analysis, the model reproduced the measured device characteristics very faithfully and did not require any parameter tuning or iterative refinement. This instant high-precision performance represents a breakthrough compared to traditional simulation-dependent approaches that require iterative calculations over hours or days.
Assistant Professor Keisuke Ide, one of the principal researchers, emphasized the transformative implications of this speed increase, saying, “Our tandem neural network achieved computational acceleration of more than six orders of magnitude compared to traditional device simulation methods. This means that what once took days can now be performed almost instantly, fundamentally changing the landscape of semiconductor diagnostics and research.”
The impact of this progress extends far beyond academic curiosity. In industrial production, instantaneous, ultra-accurate characterization of transistor material properties serves as an essential part of quality control, allowing defects and performance variations to be detected and addressed in real-time on the production line. This rapid feedback loop can significantly reduce waste and improve yields in semiconductor manufacturing facilities.
Additionally, the TNN framework paves the way for autonomous laboratories where artificial intelligence designs, executes, and interprets experiments with minimal human oversight. By quickly and reliably unraveling complex material-property relationships, AI-driven research tools can accelerate innovation cycles and optimize device designs and material formulations in ways previously unattainable due to computational bottlenecks.
Interestingly, tandem neural network architectures are not conceptually limited to semiconductors only. The researchers highlight its potential applicability to a wide range of inverse problems characterized by multivalued properties, spanning materials science, chemistry, optics, and even biological systems where multiple underlying parameter sets can produce indistinguishable observational data.
Tokyo University of Science was newly established in 2024 through the merger of Tokyo Medical and Dental University and Tokyo Institute of Technology as a base for cutting-edge interdisciplinary research. This project embodies the Institute’s mission to advance scientific understanding and technology in ways that create tangible value for society.
As the demand for faster, smaller, and more efficient semiconductor components increases, new computational tools that can quickly and accurately resolve the complex interactions of material parameters are essential. The tandem neural network approach is an important milestone in this effort, ushering in a new era in which machine learning not only complements, but fundamentally transforms research in the physical sciences.
This rise of AI-powered physical reasoning will help accelerate materials discovery, optimize manufacturing, and ultimately power the electronic devices that will drive tomorrow’s technology landscape. Using this tandem neural network, researchers have cracked open the door to solving multivalued inverse problems, once considered out of reach, at breakneck speed.
Research theme: Characterization of semiconductor materials using machine learning
Article title: Tandem neural networks quickly solve multivalued inverse problems: Application to characterization of oxide semiconductors
News publication date: May 27, 2026
Web reference: https://doi.org/10.1002/aisy.70437
image credits: Tokyo University of Science
keyword
Artificial intelligence, machine learning, semiconductor analysis, inverse problem, multivalued nature, twin neural network, material characterization, oxide semiconductor, transistor diagnosis, amorphous indium gallium zinc oxide, calculation acceleration, autonomous laboratory
Tags: Accelerated Transistor Parameter InferenceAdvanced Semiconductor Diagnostic TechnologiesCharge Carrier Mobility PredictionDevicesMachine Learning in PhysicsMachine LearningSemiconductor AnalysisMultiple Inverse Problems in SemiconductorsOvercoming Semiconductor Data DegeneracyNeural Networks for Semiconductor ModelingSemiconductor Defect Density EstimationSemiconductor Material Parameter ExtractionTandem Neural Networks for Inverse ProblemsTransistor Current Voltage Characterization
