Skip to main content

Inverse Design

Version

Inverse design capabilities are planned for MetoSim V4 (Q4 2026). This page describes the architecture and approach.

Overview​

Traditional photonics design is forward: an engineer proposes a geometry, simulates it, and evaluates performance. Inverse design flips this — you specify the desired optical response, and an optimization algorithm discovers the geometry that produces it.

MetoSim V4 will provide:

  • Adjoint gradient computation — efficient sensitivity analysis of the optical response with respect to every design pixel, in just 2 simulations (forward + adjoint)
  • Topology optimization — free-form structure discovery without pre-defined geometry parameterisation
  • AI-guided search — surrogate models (trained on V3 datasets) accelerate the design space exploration
  • Fabrication-aware constraints — minimum feature sizes, connectivity, and material compatibility enforced during optimization

How Adjoint Optimization Works​

1. Define target response       (e.g. 90% transmission at λ=1550nm)
2. Forward simulation           (standard FDTD)
3. Compute loss function        (how far from target?)
4. Adjoint simulation           (backward propagation of loss gradient)
5. Compute gradient             (∂loss/∂ε for every pixel)
6. Update geometry              (gradient descent step)
7. Repeat 2-6                   (until converged)

The key insight: step 4-5 costs only one additional simulation regardless of how many design variables exist — enabling optimization of millions of pixels simultaneously.

Architecture​

Inverse design runs as iterative Celery task chains on the existing MetoSim infrastructure:

SDK (optimization loop)
  ↓
API (submit chain)
  ↓
Redis (task chain: forward → adjoint → gradient → update → forward → ...)
  ↓
GPU Engine (FDTD forward + adjoint)
  ↓
S3 (intermediate results)

The V1→V4 architecture was designed for this from day one — iterative task chains require no re-architecture.

Planned SDK Interface​

import metosim

# Define target optical response
target = metosim.OptimizationTarget(
    monitor="transmission",
    wavelength=1.55e-6,
    target_value=0.95,        # 95% transmission
    objective="maximize",
)

# Define design region
design_region = metosim.DesignRegion(
    center=(0, 0, 0),
    size=(4e-6, 4e-6, 0.22e-6),
    materials=("Air", "Si"),   # binary optimization
    min_feature_size=100e-9,   # fab constraint
)

# Run inverse design
result = metosim.inverse_design(
    target=target,
    design_region=design_region,
    domain_size=(8e-6, 8e-6, 4e-6),
    max_iterations=100,
    learning_rate=0.01,
)

# Export optimized structure
result.export_gds("optimized_device.gds")
result.export_metofab()  # Direct to fabrication

Relationship to V3 (ML Datasets)​

V3's dataset generation capability feeds directly into V4:

  1. V3 generates thousands of forward simulations across parameter space
  2. Surrogate neural networks are trained on this data
  3. V4 uses surrogates for fast initial exploration, then refines with exact adjoint FDTD
  4. This hybrid approach reduces GPU cost by 10-100× compared to pure adjoint optimization

Research Applications​

  • Wavelength-selective metasurfaces
  • Broadband anti-reflection coatings
  • Mode converters and multiplexers
  • Metalens phase profiles
  • Photonic crystal cavity design
  • Optical computing elements