Machine Learning in Physical Design (AI-EDA) is the application of neural networks, reinforcement learning, and other ML techniques to accelerate and improve placement, routing, floorplanning, and timing optimization in chip physical design — addressing the exponential growth in design complexity that has outpaced the ability of classical algorithms to find optimal solutions within practical runtimes. ML-EDA tools have demonstrated 10–25% PPA improvement in placement and routing while reducing computational runtime, marking a fundamental shift in how electronic design automation is performed.

Why ML Is Transformative for EDA

• Classical P&R: Heuristic algorithms (simulated annealing, min-cut partitioning) → good but not optimal.
• Modern designs: Billion-transistor SoCs with 100M+ cells → search space too vast for exhaustive methods.
• ML advantage: Learn patterns from thousands of prior designs → generalize to new design problems faster.
• Key insight: Physical design has rich historical data (prior chip layouts, timing results) → ideal for supervised and reinforcement learning.

ML Applications in Physical Design

1. Placement (Cell Placement)

• Graph Neural Network (GNN) placement: Represent netlist as a graph → GNN predicts wire length and congestion for any placement configuration → guide simulated annealing.
• Reinforcement Learning (RL) placement: Train agent to place macros → reward = wire length + congestion.
• Google AlphaChip (2023): RL-based floor-planning + placement for Google TPU → reduced turnaround time from weeks to hours while achieving human-expert-quality results.
• Commercial: Synopsys DSO.ai, Cadence Cerebrus — ML-enhanced P&R optimization.

2. Routing

• Congestion prediction: Train CNN on placed netlist features → predict routing congestion before routing → feed back to placement → avoid congested configurations.
• Layer assignment: ML model predicts which net should go on which metal layer for minimum delay.
• Via optimization: RL optimizes via insertion strategy for reliability and yield.

3. Timing Prediction

• Train model on synthesized + placed netlists → predict final post-route timing without running full STA.
• Enables 10–50× faster timing feedback during RTL optimization iterations.
• GNNs trained on netlist graphs predict setup/hold slack distribution.

4. Floorplanning

• RL for macro placement: Agent places macros one at a time → reward shaped by wirelength, congestion, timing.
• GNN encoding of design connectivity → policy network suggests macro placement.

Synopsys DSO.ai and Cadence Cerebrus

How DSO.ai Works

1. Define design objectives: target timing (frequency), power, area budget
2. ML agent: Sets EDA tool options (effort levels, strategies)
3. Run EDA tools with those options → observe PPA result
4. RL feedback: Reward = how close result is to target → update policy
5. Next iteration: Agent tries different tool options guided by learned policy
6. After 50–200 iterations: Converges to near-optimal tool settings

Limitations and Challenges

• Generalization: Model trained on design A may not generalize perfectly to very different design B → requires re-training.
• Data requirements: Need thousands of prior design runs to train robust models → available only at large chip companies.
• Interpretability: RL black-box decisions hard to debug → difficult to diagnose why a particular placement was chosen.
• Integration: ML tools must plug into existing EDA flows → requires clean APIs.

Machine learning in physical design is at the inflection point of transforming EDA from human-guided heuristics to data-driven optimization — as AI-EDA tools demonstrate consistent PPA improvements and faster closure on production-quality designs, they are shifting the role of physical design engineers from manual algorithm tuning to design objective specification, promising to enable chip complexity that would be impossible to manage with classical EDA approaches alone.