🧠 Scientific vs Brute-Force Agent: AI Reasoning Under Physical Constraints

This interactive demo introduces a novel AI agent architecture that can operate under two distinct cognitive modes:
a brute-force combinatorial searcher, and a scientifically grounded reasoning agent.

This agent is capable of mode-switching, adapting its strategy to the user’s expectations:

  • In Brute-Force mode, the agent mechanically tries every possible permutation of material order and thickness combinations.
  • In Scientist mode, the agent reasons from first principles of optics, guided by spectral shape and error metrics.

🤖 Objective: Discover the correct 4-layer material order and uniform thickness

Given a target transmission spectrum, the agent must:

  • Identify the correct permutation of materials from a fixed set
  • Select a uniform thickness (10nm or 100nm)
  • Stop once cosine_similarity > 0.999999 is achieved

Constraints:

  • Materials: Si, Si₃N₄, SiO₂, AlN (use each exactly once)
  • Uniform thickness: All layers must be either 10nm or 100nm

🧬 What's New & Innovative

🔀 Mode Toggle:

  • You control whether the agent uses brute-force enumeration or scientific deduction.

🧠 Scientist Mode Intelligence:

  • Hypothesis generation and simulation driven by shape + magnitude cues
  • Uses both cosine_similarity and mean_squared_error for multi-objective evaluation
  • Fewer simulations, smarter reasoning

⚙️ Brute-Force Mode Simplicity:

  • Tries all combinations of 4! material orders × 2 thicknesses (48 total)
  • Terminates once similarity threshold is met

🔍 What Happens Under the Hood

  1. A random 4-layer stack is generated (e.g., Si → SiO₂ → AlN → Si₃N₄) and simulated using RCWA.
  2. This becomes the target spectrum.
  3. Based on the selected mode:
    • Brute-Force Agent tries all fixed permutations with simulate_spectrum(...) + cosine_similarity(...).
    • Scientist Agent reasons about spectral behavior and dynamically explores promising configurations.
  4. The agent streams its reasoning live in the UI and stops once a close match is found.

🛠️ Tools Available to the Agent:

Tool Description Arguments
simulate_spectrum Simulates optical spectrum for a layer stack layer_order: List[str], thicknesses: List[float]
cosine_similarity Measures angular similarity between spectra vec1, vec2: Lists of floats
mean_squared_error Measures magnitude mismatch in spectra vec1, vec2: Lists of floats

📊 Visualization: Below are the set of transmission spectra the agent has to search over

📊 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
Agent Mode