🧠 Multi-Agent Thin-Film Stack Optimizer (10nm + 100nm RCWA)

This interactive demo showcases a multi-agent AI system that cooperatively solves an inverse optics problem using physics-based simulation.
Instead of relying on a single agent or a single simulator, the Coordinator Agent orchestrates two specialized agents, each tuned to simulate optical spectra for different layer thicknesses (10nm and 100nm), and uses a comparison tool to evaluate results.

🤖 Objective: Discover the correct layer order and thickness

Given only a target transmission spectrum, the multi-agent system must identify:

  • The correct material permutation (from a fixed set), and
  • The correct thickness choice (10nm or 100nm, uniformly applied)

Constraints:

  • Materials: Si, Si₃N₄, SiO₂, AlN (used once each)
  • Layer thickness options: 10nm or 100nm
  • Terminate when cosine_similarity > 0.999

🧬 Agent Architecture Overview

  • 🧠 Coordinator Agent (CodeAgent | gpt-4.1-mini):
    Receives the target spectrum and is responsible for reasoning, selecting candidates, and delegating simulations to sub-agents.

  • 🔧 agent_10nm_simulator:
    Can simulate any material order using 10nm thick layers via RCWA.

  • 🔧 agent_100nm_simulator:
    Can simulate any material order using 100nm thick layers via RCWA.

  • 📏 cosine_similarity tool:
    Measures how close a simulated spectrum is to the target.

🔍 What’s Happening Under the Hood

  1. A random 4-layer stack is selected from Si, Si₃N₄, SiO₂, AlN, and simulated (at 100nm) to serve as the target spectrum.
  2. The Coordinator Agent is provided the target and access to two sub-agents:
    • agent_10nm_simulator for 10nm stacks
    • agent_100nm_simulator for 100nm stacks
  3. It explores permutations + thickness combinations by:
    • Calling a simulation agent
    • Comparing simulated output with the target using cosine_similarity
    • Continuing exploration until similarity exceeds threshold
  4. The system halts automatically once the optimal stack is found, and reports:
    • The matched material order
    • Thickness value
    • Number of permutations tried

📊 Visualization

📊 Transmission Spectra of Layer Stack Designs for all 24 permutations of material order for thickness 100nm

📊 Transmission Spectra of Layer Stack Designs for all 24 permutations of material order for thickness 10nm

🤖 Multi-Agent System Overview

🧠 Coordinator Agent: CodeAgent | openai/gpt-4.1-mini

  • ✅ Authorized Imports: ['numpy']
🛠️ Tools:
Tool Description Arguments
cosine_similarity Computes the cosine similarity between two vectors vec1, vec2: Lists of floats

🔧 Managed Agent: agent_10nm_simulator | CodeAgent | openai/gpt-4.1-mini

  • ✅ Authorized Imports: []
  • 📝 Description: Simulates optical spectra for 10nm thickness.
🛠️ Tools:
Tool Description Arguments
simulate_spectrum_10nm Simulates spectrum for 10nm layers. layer_order: List of materials

🔧 Managed Agent: agent_100nm_simulator | CodeAgent | openai/gpt-4.1-mini

  • ✅ Authorized Imports: []
  • 📝 Description: Simulates optical spectra for 100nm thickness.
🛠️ Tools:
Tool Description Arguments
simulate_spectrum_100nm Simulates spectrum for 100nm layers. layer_order: List of materials