strelka-fdtd 0.1.0

Fastest open-source differentiable photonic optimizer. JAX-accelerated 2D FDTD with free adjoint gradients.

Strelka

Fastest open-source differentiable photonic optimizer.

JAX-accelerated 2D FDTD with free adjoint gradients. Designed for inverse design of photonic neural network chips.

Key Numbers

Metric	Value
Speed vs Meep	85× faster (single-threaded)
Gradient method	jax.grad — free adjoint, no separate sim
Largest demonstrated	64×64 matrix (240K DOF, 28 min on M2)
Platform	Apple Silicon, x86, any JAX backend
License	Apache 2.0

Install

pip install strelka-fdtd

Quick Start

Python API

from strelka import optimize, simulate

# Design an 8-port photonic matrix multiplier
result = optimize(n_ports=8, design_um=14.0, resolution=10)
# → Topology optimized in ~10 seconds

# Simulate a topology
sim = simulate(result['rho'], n_ports=8)
print(f"Output power: {sim['total_power']:.4f}")

CLI

# Optimize
strelka optimize --ports 8 --res 10

# Full 64×64 (takes ~30 min on M2)
strelka optimize --ports 64 --res 5 --iters 50

# Benchmark your machine
strelka benchmark

# Decompose a weight matrix into photonic blocks
strelka decompose weights.npy --block-size 8

Block Clements Cascade

Decompose any N×N matrix into 8×8 monolithic photonic blocks:

from strelka.cascade import GivensBlockCascade
import numpy as np

W = np.random.randn(64, 64)  # your weight matrix
gbc = GivensBlockCascade(N=64, block_size=8)
result = gbc.inference(W)

print(f"Reconstruction error: {result['reconstruction_error']:.2e}")
print(f"Devices: {result['total_devices']} (vs {result['clements_mzis']} Clements MZIs)")
# → error: ~1e-15, 4× fewer devices

How It Works

1. FDTD Simulation: Full-wave electromagnetic simulation of 2D photonic structures
2. JAX Autodiff: jax.grad computes the gradient of transmission w.r.t. every pixel — no separate adjoint simulation
3. Adam Optimizer: Gradient ascent with sigmoid binarization schedule
4. Block Clements: Exact decomposition of large matrices into small monolithic blocks

Architecture

Your weight matrix (64×64)
    ↓ SVD
U (64×64 unitary) + Σ (diagonal) + V* (64×64 unitary)
    ↓ GivensBlockCascade
8×8 monolithic blocks + cross-block couplers
    ↓ strelka optimize
Foundry-ready GDS for each block

Benchmarks

On Apple M2 (8-core, 16GB):

Design	Grid	DOF	Time/iter	Full optimization
2-port	124×124	3,600	0.6s	6s
8-port	196×196	19,600	0.4s	8s
64-port	538×538	240,100	0.9s	28 min

Citing

If you use Strelka in your research, please cite:

Pushinka, "strelka-fdtd: Fastest open-source differentiable photonic optimizer," 2026.
https://github.com/venticedcappuccino/strelka-fdtd

Contributing

We welcome contributions. See CONTRIBUTING.md for guidelines.

Founding contributors receive named credit and priority access to paid work.

License

Apache 2.0