Keysight targets faster PDK development with machine learning toolkit

Keysight Technologies has introduced a new Machine Learning Toolkit aimed at significantly accelerating device modeling and Process Design Kit (PDK) development. The toolkit is part of the latest release of the Keysight Device Modeling Software Suite and is designed to cut model development and parameter extraction times from weeks to hours.

For eeNews Europe readers working on advanced semiconductor nodes, RF, or power devices, the announcement is particularly relevant as modeling complexity continues to rise while design schedules shrink. According to the release, the new AI/ML-driven approach is intended to improve productivity, model quality, and predictability across a wide range of technologies.

Addressing growing modeling complexity

The semiconductor industry is in the midst of rapid change, driven by new device architectures such as gate-all-around (GAA) transistors, the adoption of wide-bandgap materials like GaN and SiC, and the use of heterogeneous integration approaches including chiplets and 3D stacking. While these developments promise higher performance and efficiency, they also introduce major challenges for device modeling and parameter extraction.

Traditional compact modeling workflows are largely physics-based and rely heavily on manual tuning. Engineers may need to adjust hundreds of interdependent parameters across multiple operating conditions, a process that can take weeks and still fail to deliver optimal results. According to the release, this approach is increasingly unsustainable as time-to-market pressures intensify.

Keysight’s new Machine Learning Toolkit, integrated into Device Modeling MBP 2026, is designed to address these issues by combining advanced neural network architectures with ML-based optimization techniques. The toolkit includes an ML optimizer, automated extraction flows, and supporting utilities that dramatically reduce the number of required extraction steps.

From hundreds of steps to single-run optimization

Using the new toolkit, parameter extraction steps can be reduced from more than 200 to fewer than 10, according to Keysight. In practice, this means that 80 or more model parameters can be globally optimized in a single run, while still capturing secondary effects, temperature dependence, and dynamic behavior across DC, RF, and large-signal domains.

The workflow is fully automated and integrates with Keysight’s existing Device Modeling platform, with support for Python-based customization. The company says the approach scales across multiple technologies, including FinFET, GAA, GaN, SiC, and bipolar devices, allowing reusable modeling flows across process nodes. This, in turn, enables faster Design Technology Co-Optimization (DTCO) and shortens PDK development cycles from weeks to days.

Alongside the ML toolkit, Keysight also announced updates to other device modeling products, including new aging model QA rules in Device Modeling MQA 2026 and enhanced low-frequency noise testing capabilities in WaferPro 2025 and A-LFNA 2026, further extending its device modeling portfolio.