AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications

Fahad Usmani Device Modeling Expert Keysight Technologies (US) [email protected] December 12, 2025 MOS-AK LatAm AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications Agenda •Introduction: Device Modeling at Keysight •AI/ML innovations in Device Modeling (DM) •ML Toolkit using Python, ML optimizer, and ANN toolbox •Full & Hybrid ANNs •Applications: •Parameter extraction for advanced CMOS nodes •Model development for RF GaN- & GaAsHEMTs •Modeling for SiC or IGBT power applications •AI/ML efforts at Keysight: •Major Enhancements •EDAchat •EDA Copilots •Conclusion •Q&A AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications 2 AI/ML Use-cases AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications AI for Modeling •Solving the multi-dimensional problem of parameter extraction •Improve workflow efficiency by simplifying flows •Enhance accuracy of extraction results •Complement limitations of compact models •Shorten model development cycles AI for Design •Create new designs based on objectives and constraints •Explore design space faster •Make design flows more autonomous •Optimize various circuit topologies •Solve design trade-offs in Digital/Analog/RF systems 3 How can AI/ML techniques be successfully applied in Device Modeling? AI/ML solutions at Keysight AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications 𝑉gs 𝑉ds 𝑇j 𝜑1 𝜑2 ID DIGITAL TWIN Keysight EDA Chat (LLM) Learned best tricks from 100s of Keysight Experts, Documents, resources Artificial Intelligence Machine Learning Deep Learning Generative Intelligence Autoencoders ML-driven Optimizer Full Compact Model in Verilog-A Keysight ANN Library Hybrid ANN Model in Verilog-A Modified Topology with added ANN elements Investigation Measured Data ML-driven autoextraction flows ANN Toolbox Hybrid ANN ML Optimizer U0 VTH0 ML-based Deep RL PCA ML Toolkit Surrogate Models SymPy 4 •NeuroFET: Similar to the RootFET, but with a polynomial replaced by an ANN. Intrinsic ANN consists of Train + Test splits: •Neuron Activation functions (Sigmoid/Hyperbolic tangent) •Training method (Quasi-Newton/Levenberg-Marquardt) •Number of hidden neurons •DynaFET: NQS Large-signal III-V FET model offered an enhanced solution for GaN-HEMTs •Intrinsic model (ANN with thermal and trapping effects) + parasitic R, L, C •IV, S-par fitting done on large-signal waveform data from NVNA. NVNA Waveforms Industry’s First AI/ML Solutions in Device Modeling NeuroFET and DynaFET models AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications ANN ANN + SHE DC, S-par, Large-signal Waveforms with de-embedding Encrypted model NeuroFET DynaFET NVNA 5 ANN Toolbox: GaAs-HEMT Keysight ANN Model Generator in IC-CAP AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications •Useful for quick point model generation at any design phase or when accurate models are unavailable •Keysight ANN implementation generates ANN from partial derivative data and outputs Verilog-A SPICE model •Consists of Python functions to •Train ANNs per input data •Evaluate ANN model •Output ANN eval circuit for simulation •GaAs-HEMT as an example: 4 sub-ckt ANN models for Id/Ig/Qd/Qg combined •Not used for replacing scalable compact models IV results S-Parameters Data ANN Model ANN Toolbox: Cryogenic CMOS modeling Enabling models and circuits for Quantum Computing AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications observed but never previously simulated Data Model P.A. ‘t Hart, J. Xu, J. van Staveren, M. Babaie, D.E. Root, and F. Sebastiano, “Artificial Neural Network Modelling for Cryo-CMOS Devices”, IEEE 14th Workshop on Low Temperature Electronics (WOLTE), April 2021. •Advanced node CMOS devices exhibit altered characteristics at very low temperatures •Existing compact models may not fully capture freeze-out effects across all bias/temp ranges •Extraction procedures are not straightforward 7 Hybrid ANN solution: GaN-HEMT devices •Hybrid ANN modeling approach compensates for unmodeled non-linear physical behavior by incorporating additional R/C elements ✓Retain the physics-based core model and its ability to provide insights ✓Improves the overall accuracy of the model ×Requires expert investigation and validation. •ASM-HEMT Cgd vs. Vds model fitting improved in ASM-HEMT by applying Hybrid-ANN AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications Problem: How to overcome the limitations of physics-based compact models? C-V I-V R. P. Martinez et al., "A Hybrid Physical ASM-HEMT Model Using a Neural Network Based Methodology," 2024 IEEE BiCMOS and Compound Semiconductor Integrated Circuits and Technology Symposium (BCICTS), Fort Lauderdale, FL, USA, 2024, pp. 3841,doi: 10.1109/BCICTS59662.2024.10745660 8 Hybrid ANN solution: IGBT Double Pulse Test (DPT) •Model may fit in some bias conditions, but RR model itself is not bias-dependent. •Hybrid ANN approach is used to create a bias-dependent equation to model RR of body diode @multibias → •An empirical equation is used to extract the on-state capacitance ↓ AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications Challenge: No bias-dependent model for reverse-recovery (RR) and lack of on-state cap model Shih, A., et al. "A Double Pulse Test Based IGBT On-State Capacitance Extrac-tion of ANN-assisted Hybrid Model." PCIM Conference 2025; International Exhibition and Conference for Power Electronics, Intelligent Motion, Renewable Energy and Energy Management. VDE, 2025. •Hybrid ANN models can be used to adjust the accuracy of encrypted models supplied by foundries •ANN can overfit or underfit depending on its configuration (layers/neurons) DPT transient simulations DPT circuit 9 DM Product Portfolio AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications Services Services WaferPro Device Modeling & Characterization •Data analysis •Model extraction •Model verification •Automated measurements (1/f, Load-pull, S-par, IV/CV) Device Modeling Services •Modeling •PDK creation •Reference design flows •Front-to-back PDK support 17 Resources •What’s New in Device Modeling webpage: https://www.keysight.com/us/en/lib/resourc es/software-releases/whats-new-in-devicemodeling.html •Recent blogs: •https://semiengineering.com/ai-meets-device-modeling-transforming-compact-modeling-with-machine-learning/ •https://www.keysight.com/blogs/en/tech/sim-des/from-equations-to-intelligence •Products: •W7009E: IC-CAP ANN modeling toolkit •W7301B: IC-CAP Model Generator Plus •W7300B: IC-CAP Device Modeling Platform •W6023E: MBP Machine Learning Toolkit •W6329B: MBP core, QA Express, and Machine Learning toolkit AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications 18 Q & A AI/ML-Driven Device Modeling for Advanced Nodes, RF and Power Applications 19