UNCHARTED: A Universal ML-based Framework for Circuit Simulation and Inverse Design

Professor Wanyeong Jung’s group at KAIST has developed UNCHARTED, a machine learning framework that learns the physical behavior of individual circuit components and uses this knowledge for both circuit simulation and automatic circuit optimization. Unlike many existing AI-assisted circuit design approaches that treat circuits as black boxes, UNCHARTED models transistors, passive components, and electrical nodes individually. This leads to a finer-grained understanding of signal propagation across different circuit topologies. The work was presented at the 63rd ACM/IEEE Design Automation Conference (DAC 2026).

UNCHARTED represents circuits as bipartite graphs, explicitly separating electrical nodes from devices such as transistors, resistors, and capacitors. A spatio-temporal graph neural network then propagates information through both circuit topology and time. Using an auto-regressive structure with external driving signals, the framework predicts node voltages and currents over time, effectively emulating transient simulation. Physical parameters such as transistor dimensions, resistance, and capacitance are incorporated through a parameter encoder inspired by low-rank adaptation (LoRA).

Experiments showed that UNCHARTED can substantially accelerate circuit simulation while maintaining waveform accuracy. For CMOS circuits, more than 70% of test cases achieved over 95% waveform similarity to SPICE simulation. Compared with Cadence Spectre APS, the framework achieved simulation speedups ranging from 67.6× to 241.9×, demonstrating its potential as a fast neural alternative for transient circuit analysis.

Beyond simulation, UNCHARTED can also serve as a differentiable circuit optimizer. By freezing the trained neural model and back-propagating design objectives to physical parameters, transistor dimensions and passive-component values can be optimized directly. In a six-stage delay-chain example, the optimized design achieved a SPICE-validated delay of approximately 0.302–0.303 ns, close to the theoretical optimum of 0.300 ns. The framework was also applied to a differential amplifier, where it optimized device parameters for multiple target gains while considering power consumption. By combining component-level modeling, neural simulation, and inverse design in a single framework, UNCHARTED provides a foundation for future AI-assisted electronic design automation systems with applications such as circuit sizing, PVT analysis, and technology transfer.

This research was supported by Samsung Electronics Co., Ltd., the Ministry of Science and ICT (MSIT) of the Korean government, the Institute of Information & Communications Technology Planning & Evaluation (IITP), the National Research Foundation of Korea (NRF), and the AI Computing Infrastructure Enhancement Program.