Breakthrough in Semiconductor Testing Efficiency
Researchers have developed a novel machine learning approach that dramatically speeds up electrical characterization of advanced semiconductor devices. The work, published in the October 2026 issue of Engineering Applications of Artificial Intelligence, introduces physics-informed Gaussian Process Regression combined with active sampling and warm-starting techniques. This method reduces the number of required measurements by a factor of six while preserving high accuracy in key parameters such as threshold voltage and subthreshold slope.
The study was led by Husnu Murat Kocak of IMEC and KU Leuven, with co-authors Jerome Mitard from IMEC, Lorenzo Perini, Ahmet Teoman Naskali, and Jesse Davis from KU Leuven. Their findings address a critical bottleneck in semiconductor development where traditional characterization methods struggle to keep pace with increasingly complex chip designs.
Understanding Electrical Characterization in Modern Semiconductors
Electrical characterization involves measuring how transistors respond to different voltages, producing characteristic curves that reveal performance metrics. For complementary field-effect transistor (CFET) architectures used in cutting-edge chips, this process typically requires sweeping gate voltages across many points. Each measurement takes time, and with billions of transistors per wafer, the cumulative effort becomes substantial.
Process variations from lithography, temperature, and alignment introduce subtle differences even among devices of the same design. Capturing these variations accurately demands dense sampling, yet excessive measurements inflate costs and extend development timelines. The new research targets this trade-off directly by leveraging prior knowledge from previously tested devices.
The Role of Gaussian Processes in Data-Efficient Modeling
Gaussian Process Regression (GPR) provides a probabilistic framework for modeling functions with uncertainty estimates. In this context, it predicts drain current values across the full voltage range based on a small number of actual measurements. Active sampling then selects the next most informative voltage points by focusing on regions of high model uncertainty.
Traditional “cold start” GPR begins with no prior information for each new device. The innovation here lies in warm-starting the model using curves from similar previously characterized devices. Similarity is determined through physics-informed features such as channel thickness, doping profiles, and design parameters, allowing the system to initialize with a realistic reference curve rather than a generic mean function.
Physics-Informed Warm Starting and Active Sampling
The team evaluated multiple strategies for selecting reference curves and integrating them into the GPR framework. By anchoring the model in physically plausible prior knowledge, the approach converges faster and requires fewer measurements to achieve target accuracy. Experiments across 1,000 devices spanning varied designs, mask sets, and process conditions demonstrated consistent gains over both conventional fixed-grid sampling and cold-start GPR.
Key performance metrics included threshold voltage accuracy within a maximum absolute error of 25 millivolts and subthreshold slope within 25 millivolts per decade. These tolerances align with industry requirements for reliable device qualification while slashing characterization time.
Photo by Frantzou Fleurine on Unsplash
Empirical Validation Across Diverse Device Populations
Testing encompassed complementary field-effect transistors fabricated under different process conditions at IMEC facilities. The methodology proved robust to variations in device geometry and material properties. Warm-started models consistently outperformed baselines, with the largest improvements observed when a sufficient pool of prior devices was available for reference selection.
The study also explored how the number of available prior devices influences performance. Even modest reference pools yielded meaningful reductions in measurement count, highlighting the practical value of accumulating characterization data across projects.
Broader Implications for Semiconductor Research and Industry
Accelerating characterization directly shortens design cycles and lowers costs in an industry facing relentless demand for faster, more efficient chips. The approach is generalizable beyond semiconductors to any signal-acquisition task where prior measurements exist and physics constrains the underlying relationships.
Academic and industrial labs can integrate the technique into existing test equipment with relatively modest software additions. The work builds on an earlier 2025 study by the same lead author that introduced active sampling without warm starting, showing iterative refinement of AI methods in real-world engineering contexts.
View the original publication here: https://www.sciencedirect.com/science/article/abs/pii/S0952197626017823.
Connections to Machine Learning in Engineering Education
The research exemplifies how probabilistic modeling and active learning techniques translate from theory to high-stakes manufacturing environments. University programs in electrical engineering, materials science, and computer science can draw on this case study to illustrate the integration of physics-informed machine learning with domain-specific constraints.
Graduate students and postdoctoral researchers working on semiconductor characterization or similar measurement-intensive processes now have a concrete demonstration of how warm-starting strategies reduce experimental burden without sacrificing fidelity.
Future Directions and Open Questions
Potential extensions include scaling the similarity metrics to larger device libraries, incorporating multi-fidelity data sources, and adapting the framework to emerging transistor architectures. Researchers may also explore transfer learning across different fabrication facilities or technology nodes.
The methodology’s emphasis on uncertainty quantification aligns well with ongoing efforts to make AI systems more interpretable and trustworthy in regulated industries such as semiconductors.
Stakeholder Perspectives from Research and Manufacturing
Device engineers benefit from faster turnaround on qualification data, enabling quicker iteration on process tweaks. AI researchers gain a high-impact application domain that rewards careful integration of domain knowledge. Funding agencies and industry consortia may view such efficiency gains as strategic advantages in global semiconductor competitiveness.
IMEC’s involvement underscores the value of collaborative environments where academic machine learning expertise meets industrial fabrication capabilities.
Actionable Insights for Research Teams
Labs interested in adopting similar methods should begin by cataloging historical characterization curves with associated device metadata. Implementing a similarity search module and integrating it with an active sampling loop can yield immediate efficiency improvements. Open-source Gaussian process libraries provide a practical starting point for prototyping.
Training programs that combine semiconductor physics with modern probabilistic modeling will prepare the next generation of researchers to tackle measurement bottlenecks across engineering disciplines.