Institute of Science Tokyo Develops ML Framework to Infer Semiconductor Material Parameters with High Accuracy in Under 1 Millisecond from Transistor Measurements

June 28, 2026

Institute of Science Tokyo / Advanced Intelligent Systems Japan

Overview

Researchers at the Institute of Science Tokyo have developed a groundbreaking machine learning framework for solving inverse problems in semiconductor materials. This tandem neural network enables the inference of physical parameters of semiconductor materials from transistor measurements with high accuracy in under one millisecond. It offers a dramatic speedup over conventional iterative optimization methods, with applications in manufacturing quality checks and real-time analysis for autonomous laboratory systems, poised to significantly impact the semiconductor industry.

In Depth

Key Findings

A research team at the Institute of Science Tokyo has developed a groundbreaking machine learning framework capable of solving complex ‘inverse problems’ in semiconductor material analysis with unprecedented speed and accuracy. This tandem neural network allows for the inference of physical parameters of semiconductor materials from transistor measurements in less than one millisecond, all while maintaining high precision. This technology holds the potential to dramatically accelerate quality control and material development processes in semiconductor manufacturing.

Technical / Clinical Details

The framework employs a ‘tandem structure,’ arranging two neural networks in series. The first network functions as a ‘forward model,’ predicting the electrical characteristics of a transistor (e.g., current-voltage curves) from the material’s physical parameters (e.g., bandgap, mobility, impurity concentration). The second network, trained using data generated by this forward model and actual experimental data, acts as an ‘inverse model,’ inferring the original physical parameters from the measured transistor characteristics. Once trained, this inverse model can directly infer material parameters from new measurement data without requiring iterative numerical optimization. Conventional inverse problem-solving methods, based on nonlinear optimization algorithms, typically required seconds to minutes to obtain a solution. This new framework reduces that to less than one millisecond while significantly improving the accuracy of parameter inference.

Background & Context

The semiconductor industry, exemplified by Moore’s Law, constantly demands miniaturization, higher performance, and lower costs. When developing new semiconductor materials and device structures, accurately identifying their physical parameters is crucial for performance evaluation, yield improvement, and fault diagnosis. However, estimating these internal parameters from electrical measurements (‘inverse problems’) has been known to be mathematically challenging and computationally expensive. This bottleneck has slowed down the R&D cycle for new materials and hindered the optimization of manufacturing processes. The introduction of AI, particularly deep learning, is gaining traction as a powerful means to overcome this challenge, with applications in the semiconductor sector advancing as part of materials informatics.

Strategic Significance & Outlook

This machine learning framework developed by the Institute of Science Tokyo is poised to have a profound impact on the semiconductor industry. For real-time quality control on manufacturing lines, rapid parameter estimation at the wafer level will enable early detection of defects and immediate process adjustments, leading to dramatic yield improvements. In autonomous laboratory systems, integrating this technology into AI agents’ decision-making processes will further accelerate and autonomize the entire cycle from material exploration to device fabrication and evaluation. In the future, this versatile inverse problem-solving framework is expected to be applied not only to the semiconductor field but also to the characterization and design of other functional materials, such as quantum materials, battery materials, and catalysts, thereby collectively enhancing the speed and efficiency of scientific discovery.

Source: https://www.isct.ac.jp/en/news/tky57fxj4rub

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.