AI-assisted device modeling accelerates PDK development

Keysight Technologies has released a new Machine Learning Toolkit as part of its latest Device Modeling Software Suite, aiming to significantly reduce the time required for compact model development and parameter extraction in advanced semiconductor processes.

Increasing modeling complexity in advanced nodes
The transition toward gate-all-around (GAA) transistors, wide-bandgap materials such as gallium nitride and silicon carbide, and heterogeneous integration approaches including chiplets and 3D stacking has increased the complexity of device modeling. Conventional physics-based compact models typically require engineers to manually tune hundreds of interdependent parameters across DC, RF, temperature and large-signal operating conditions. This process can extend over several weeks and may still leave gaps in predictive accuracy.

As process design kits (PDKs) are expected to be delivered earlier in the technology lifecycle, these manual workflows have become a bottleneck for both foundries and integrated device manufacturers.

Machine learning–driven parameter extraction
The new Machine Learning Toolkit, integrated into Device Modeling MBP 2026, introduces a framework that combines neural-network architectures with ML-based optimization. According to Keysight, automated extraction flows can reduce parameter-extraction steps from more than 200 to fewer than 10, enabling global optimization of more than 80 parameters in a single run.

This approach is intended to capture secondary effects, temperature dependencies and dynamic behaviors more consistently than manual tuning, while improving agreement between measured and simulated behavior across DC, RF and large-signal domains.

Automation and scalability across technologies
The toolkit is designed to operate within Keysight’s existing device-modeling environment and supports Python-based customization. This allows modeling teams to integrate ML-driven extraction into established workflows rather than replacing them entirely.

Keysight states that the automated flows are adaptable across multiple device architectures, including FinFET, GAA, GaN, SiC and bipolar technologies. This reusability is intended to support consistent modeling practices across different process nodes and material systems.

Implications for DTCO and time-to-market
By shortening model extraction and validation cycles, the ML-based approach is positioned to improve design technology co-optimization (DTCO). Faster feedback between device and circuit design can reduce PDK development timelines from weeks to days, supporting earlier design starts and tighter alignment between process and product requirements.

Related updates in the modeling portfolio
Alongside the Machine Learning Toolkit, Keysight has introduced updates across other device-modeling tools. Device Modeling MQA 2026 adds new quality-assurance rules for aging models, while WaferPro 2025 includes remote-control capabilities for low-frequency noise testing. The A-LFNA 2026 release extends low-frequency noise stress testing, enabling continuous measurement from stress to noise characterization.

Together, these updates reflect a broader shift toward more automated, data-driven device modeling workflows as semiconductor technologies continue to scale in complexity.