Dielectric

Computational Geometry + Multi-Agent AI for PCB Design

Dielectric is an AI-powered PCB design platform that combines advanced computational geometry algorithms with multi-agent AI reasoning to generate, optimize, and analyze printed circuit board designs from natural language descriptions.

---

Overview

Dielectric provides tools for PCB design generation, optimization, and analysis. The platform uses computational geometry analysis, simulated annealing optimization, and Grok (xAI) reasoning to support:

• Generate complete PCB designs from natural language descriptions
• Optimize existing designs using AI-powered algorithms
• Analyze designs with computational geometry metrics
• Export to industry-standard formats (KiCad, Gerber, etc.)

---

Key Features

Natural Language Design Generation

Input a natural language description of your PCB design. The system generates component placement and routing suggestions based on the description.

Example:

"Design a 2.4GHz RF module with transceiver IC, matching network, antenna, and power management"

Computational Geometry Analysis

The platform implements computational geometry algorithms for PCB layout analysis:

• Voronoi Diagrams - Analyze component distribution and thermal hotspots
• Minimum Spanning Tree (MST) - Optimize trace routing and minimize wire length
• Convex Hull Analysis - Measure board utilization and component spread
• Delaunay Triangulation - Identify component neighborhoods and local density
• Thermal Hotspot Detection - Gaussian thermal diffusion modeling
• Net Crossing Analysis - Detect routing conflicts and complexity
• Force-Directed Layout Metrics - Analyze placement equilibrium

Multi-Agent AI Architecture

The system uses multiple specialized agents:

• IntentAgent - Understands user goals using computational geometry + xAI
• DesignGeneratorAgent - Creates designs from natural language
• LocalPlacerAgent - Optimizes component placement (deterministic, fast)
• GlobalOptimizerAgent - Applies research-backed global optimization algorithms
• VerifierAgent - Validates design rules and constraints
• ErrorFixerAgent - Automatically fixes violations
• RoutingAgent - Handles trace routing and differential pairs
• ManufacturingAgent - Generates Gerber, drill, and pick-place files

Comprehensive Design Analysis

Compare designs before and after optimization with:

• PCB Layout Comparison - Visualize component placement improvements
• Schematic Comparison - See net routing optimizations
• Thermal Analysis - Identify and resolve thermal hotspots

Visual Design Comparison

The platform provides before/after visualizations for optimization analysis:

PCB Layout Comparison

Before and After PCB Layout Comparison showing optimized component placement and routing. The optimized layout demonstrates shorter trace lengths, reduced crossings, and improved component organization.

Schematic Comparison

Schematic Comparison demonstrating improved net routing and shorter trace lengths. Color-coded nets (VCC, GND, XTAL1, RESET) show optimized connectivity patterns.

Thermal Analysis

Thermal Analysis showing reduced hotspot concentration after optimization. The heatmap visualization demonstrates improved thermal distribution with less intense hotspots.

Advanced Optimization

The platform implements multiple optimization strategies:

• Enhanced Simulated Annealing - With xAI reasoning integration
• Force-Directed Placement (FDP) - Based on research literature
• Quadratic Placement - For large-scale designs
• Analytical Placement with Partitioning - Hierarchical optimization
• Compaction Algorithms - Space-efficient placement

Industry-Standard Export

Export your optimized designs to:

• KiCad - Full .kicad_pcb file format
• Gerber Files - Manufacturing-ready output
• Drill Files - Excellon format
• Pick-and-Place Files - CSV/JSON for assembly
• BOM (Bill of Materials) - Component lists with pricing

KiCad Export

Exported design opened in KiCad PCB Editor, showing components, pads, and ratsnest connections. The design is ready for further editing, routing, or direct manufacturing preparation.

---

Architecture

Computational Geometry Engine

Dielectric's geometry analyzer computes real-time metrics that feed into AI reasoning:

{
    "density": 0.003,  # components/mm²
    "voronoi_variance": 0.45,  # distribution uniformity
    "mst_length": 234.5,  # mm - trace length optimization
    "convex_hull_area": 8500.0,  # mm² - board utilization
    "thermal_hotspots": 3,  # number of hotspot regions
    "thermal_risk_score": 0.65,  # 0-1 thermal risk
    "net_crossings": 12,  # routing conflicts
    "routing_complexity": 0.42,  # 0-1 complexity score
    "overlap_risk": 0.08,  # collision risk
    "force_equilibrium_score": 0.89  # placement stability
}

AI Reasoning Integration

Grok (xAI) reasoning is integrated at multiple stages:

1. Design Generation - Extensive reasoning about component selection and placement
2. Intent Analysis - Maps user intent to optimization weights using geometry data
3. Optimization Strategy - Real-time reasoning during simulated annealing (every 25 iterations)
4. Post-Optimization Analysis - Provides insights and refinement suggestions

Multi-Agent Workflow

User Intent → IntentAgent → Geometry Analysis → xAI Reasoning
                                    ↓
                    ┌───────────────┴───────────────┐
                    ↓                               ↓
        DesignGeneratorAgent          LocalPlacerAgent
                    ↓                               ↓
            GlobalOptimizerAgent      VerifierAgent
                    ↓                               ↓
            RoutingAgent              ErrorFixerAgent
                    ↓                               ↓
            ManufacturingAgent        ExporterAgent

---

Installation

Prerequisites

• Python 3.10+
• xAI API key (for Grok reasoning)
• KiCad (optional, for library integration)

Setup

# Clone the repository
git clone <repository-url>
cd dielectric

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Set up environment variables
export XAI_API_KEY=your_xai_api_key_here

Running the Application

Backend Server:

cd dielectric
source venv/bin/activate
uvicorn src.backend.api.main:app --host 0.0.0.0 --port 8000 --reload

Frontend:

cd dielectric/frontend
streamlit run app_dielectric.py

The application will be available at http://localhost:8501

---

Usage

Generate Design from Natural Language

1. Select "Generate Design" workflow
2. Enter a natural language description of your PCB design
3. Set board dimensions (width and height in mm)
4. Click "Generate Design"
5. Review the generated design with visualizations

Example Descriptions:

• "Design an audio amplifier with op-amp, input/output capacitors, and power supply filtering"
• "Create a switching power supply with buck converter IC, inductor, capacitors, and feedback resistors"
• "Design a sensor interface board with ADC, voltage reference, and signal conditioning"

Optimize Existing Design

1. Select "Optimize Design" workflow
2. Upload your PCB design file (.kicad_pcb, .json, or folder/zip)
3. Enter optimization intent (e.g., "Minimize trace length", "Reduce thermal hotspots")
4. Click "Optimize"
5. Review optimization results:
   • Before/After comparisons (Layout, Schematic, Thermal)
   • Computational geometry metrics
   • Multi-agent workflow status
6. Export optimized design to KiCad or manufacturing files

Understanding Optimization Results

The platform provides analysis including:

Geometry Metrics:

• Voronoi Variance - Lower is better (more uniform distribution)
• MST Length - Shorter is better (minimized trace routing)
• Thermal Risk Score - Lower is better (reduced hotspots)
• Routing Complexity - Lower is better (fewer conflicts)

Visual Comparisons:

• PCB Layout - Component placement optimization
• Schematic - Net routing improvements
• Thermal - Heat distribution analysis

---

API Endpoints

Design Generation

POST /generate
Content-Type: application/json

{
    "description": "Design an audio amplifier...",
    "board_size": {
        "width": 120,
        "height": 80,
        "clearance": 0.5
    }
}

Optimization

POST /optimize
Content-Type: application/json

{
    "board": {...},
    "components": [...],
    "nets": [...],
    "intent": "Minimize trace length and thermal hotspots"
}

Export

POST /export/kicad
Content-Type: application/json

{
    "placement": {...}
}

Manufacturing Files

• POST /manufacturing/gerber - Generate Gerber files
• POST /manufacturing/drill - Generate drill files
• POST /manufacturing/pick-place - Generate pick-place files
• POST /manufacturing/jlcpcb/upload - Upload to JLCPCB

Advanced Features

• POST /routing/auto - Auto-route design
• POST /routing/differential-pairs - Route differential pairs
• POST /simulation/signal-integrity/analyze - SI analysis
• POST /simulation/power-integrity/analyze - PI analysis
• POST /simulation/thermal/heatmap - Thermal heat map
• POST /bom/generate - Generate Bill of Materials
• POST /drc/advanced - Advanced Design Rule Checks

---

Technical Details

Computational Geometry Algorithms

Dielectric implements research-backed algorithms from:

• Dorneich et al. - "Global Optimization Algorithms for Chip Layout and Compaction"
• Force-Directed Placement - For balanced component distribution
• Quadratic Placement - For large-scale optimization
• Analytical Placement - With hierarchical partitioning

AI Reasoning

• Model: Grok (xAI) via API
• Reasoning Frequency: Every 25 iterations during optimization
• Context: Full computational geometry metrics + user intent
• Output: Optimization strategy, component suggestions, thermal insights

Performance

• Analysis Time: ~50-200ms per geometry analysis
• Optimization: Configurable iterations (default: 200)
• Design Generation: ~30-180 seconds (depends on complexity)
• Scalability: Handles designs with 100+ components

---

Project Structure

dielectric/
├── src/
│   ├── backend/
│   │   ├── agents/          # AI agents (Design, Intent, Optimizer, etc.)
│   │   ├── geometry/        # Computational geometry algorithms
│   │   ├── optimization/    # Simulated annealing, global optimizers
│   │   ├── routing/          # Autorouting, differential pairs
│   │   ├── manufacturing/   # Gerber, drill, pick-place generation
│   │   ├── simulation/      # SI, PI, thermal analysis
│   │   ├── ai/              # xAI client integration
│   │   └── api/             # FastAPI endpoints
│   └── frontend/
│       └── app_dielectric.py # Streamlit UI
├── docs/
│   ├── COMPUTATIONAL_GEOMETRY_USAGE.md
│   ├── PRODUCT_EXPANSION_PLAN.md
│   └── images/              # Screenshots and diagrams
└── README.md

Dielectric is built with:

• Backend: FastAPI, Python 3.10+
• Frontend: Streamlit
• AI: Grok (xAI) API
• Geometry: SciPy, NumPy
• Visualization: Plotly

Contact

abiral@princeton.edu