Agentic Design Automation

Project Overview

This is the umbrella of my Ph.D. dissertation, An Agentic System for End-to-End VLSI Physical Design. The central idea: an LLM-driven agent that plans and executes the full physical-design flow — floorplanning, placement, legalization, and routing — by invoking ML-driven engines and commercial tools under unified PPA objectives.

Key Ideas

Agent Over a Real Design Flow

The agent orchestrates the floorplan-to-route flow, deciding which engine or tool to invoke at each stage and how to trade off power, performance, and area, rather than following a fixed script.

Grounded in Reliable Primitives

The agent is grounded in core contributions it invokes as dependable building blocks:

• a TwinTree-based quality-preserving legalizer (ICCAD 2026), and
• GNN-guided design engines for timing-driven placement and floorplanning.

Staying on Physically Valid States

Agent actions are kept on physically valid states through built-in topological invariants (from the TwinTree representation) and solver-based checking of constraint violations — so autonomous exploration never drifts into illegal layouts.

Why It Matters

Modern design flows stitch together many specialized engines and commercial tools, each with its own knobs. An agent that reasons over the whole flow — while being constrained to physically valid states and unified PPA objectives — points toward more automated, adaptive, and reliable design closure.

Building Blocks

• Legalization: Physical Design Methodology & Legalization
• ML engines: ML for Physical Design

Future Directions

• Broader tool coverage across the floorplan-to-route flow.
• Learned policies for engine/tool selection under PPA budgets.
• Verifiable guarantees on agent actions via invariants and solver-based checking.