860x less power than electronic AI. 99.5% fidelity after 106 krad radiation. Athermal, no active compensation needed.
Photonic neural network chip that does AI inference using light instead of electrons. Matrix multiplication at the speed of light, 1/860th the power of an ARM Cortex-M4F, and inherently radiation-tolerant because photons don't care about cosmic rays.
Patent pending: Geometry-Protected Photonic Metamaterial Interconnects (U.S. Provisional Application No. 63/987,333).
129.5 pJ per inference vs. 111,370 pJ for an ARM Cortex-M4F. Three orders of magnitude less energy, critical when your power budget is a solar panel.
2x2 Hadamard matrix multiplier: 100x100 pixel design region on 220nm SOI. Every pixel placed by adjoint-method topology optimization against a fabrication robustness objective. Not hand-designed waveguides, computationally optimal photonic structures.
Monte Carlo radiation analysis: 90 simulations across 9 dose levels (0 to 150 krad). Our model predicts 0.02% degradation at 10 krad, consistent with NASA ISS measurements showing no degradation after 1 year of on-orbit exposure. Fidelity stays above 99% even at 150 krad.
At 15-year LEO mission duration, electronic 28nm CMOS suffers 20.6% catastrophic failure rate from single-event upsets. Our photonic chip: 0.34% graceful degradation. Photons don't flip bits.
96.7% accuracy on Iris dataset (30 test samples). Single misclassification: one Versicolor classified as Virginica, the two hardest-to-separate classes in the dataset.
Multi-material topology optimization (Si / SiO2 / Polymer) drives the effective thermo-optic coefficient from 1.86e-4 K^-1 down to < 1e-6 K^-1. 99.6% reduction, thermally stable across orbital temperature swings (-170C to +120C) without active heater compensation.
Our photonic chip sits in the bottom-left corner: lowest power, competitive radiation tolerance. Every other flight processor trades power for rad-hardness. We get both.
| Metric | Achieved |
|---|---|
| Power vs. ARM Cortex-M4F | 860x reduction (129.5 pJ vs. 111,370 pJ) |
| Classification accuracy | 96.7% (Iris), 96.94% (digits) |
| Radiation tolerance (15yr LEO) | 0.34% degradation vs. 20.6% electronic |
| SEU reliability vs. 28nm CMOS | 644x more reliable |
| ISS flight data consistency | Prediction matches 1-year on-orbit measurement |
| Athermal dn_eff/dT reduction | 99.6% (no active compensation) |
| Fabrication robustness score | 9.2 / 10 |
| Worst-case transmission (±100nm) | 0.82 |
| Insertion loss | 0.5 dB |
| Flatness ratio | 0.92 |
| 64x64 device count reduction | 5.2x vs. Clements architecture |
Electronic AI chips store information as charge on transistor gates. A single cosmic ray can flip those charges, corrupting weights or activations mid-inference. In LEO, this happens roughly once per day per chip. The standard fix is triple modular redundancy: three copies of everything, 3x the hardware, 3x the power.
Photonic chips encode information as the phase and amplitude of light propagating through silicon waveguides. There is no stored charge to flip. A cosmic ray passes through the waveguide and the light doesn't notice. This is not radiation shielding or error correction. It is a fundamental property of the physics.
| Mission Profile | Why Photonic AI Matters |
|---|---|
| Mars surface autonomy | 4-24 minute comms delay makes real-time Earth control impossible. Rover needs onboard AI for hazard avoidance and science target selection. Power budget: ~100W total for the entire rover. |
| LEO constellation edge AI | Thousands of small satellites need onboard image classification to downlink only useful data. 5-15W per satellite, 15-year lifetime, no servicing. |
| Lunar Gateway | Autonomous rendezvous and docking in cislunar space. Radiation environment is unshielded by Earth's magnetosphere. |
| Europa lander | Extreme radiation (5.4 Mrad over mission), extreme cold (100 K surface), no real-time communication. The lander must decide what to study on its own. |
| Deep space probes | Voyager-class missions where the processor must work for decades with zero maintenance at microwatt power. |
- Platform: 220nm silicon-on-insulator (SOI), C-band (1550 nm)
- Foundry compatible: AIM Photonics multi-project wafer runs (Albany, NY)
- Design region: 100x100 pixels at 30 px/um = 3.3 x 3.3 um per 2x2 block
- Morphological robustness: designs validated against ±50nm and ±100nm fabrication variation (erosion, dilation, opening, closing)
- Binary topology: final designs are fully binarized (Si or SiO2), no greyscale lithography required
The physics and validation in this work build on peer-reviewed results:
| Claim | Supporting Reference |
|---|---|
| Photonic ICs survive space radiation | Tzintzarov et al., "Space-qualifying silicon photonic modulators and circuits," Science Advances 10, eadi9171 (2024). 325 days on ISS, no measurable degradation. |
| Inverse-designed PNNs achieve high density | Kottapalli et al., "Inverse-designed nanophotonic neural network accelerators," Nature Communications (2026). 400M parameters/mm2. |
| Thermo-optic coefficient of silicon | Cocorullo et al., J. Appl. Phys. 86, 3281 (1999). dn/dT = 1.86 x 10^-4 K^-1. |
| Photoelastic tensor of silicon | Biegelsen, Phys. Rev. B 9(5), 2635 (1974). p11 = -0.094, p12 = 0.017. |
| Film stress from CTE mismatch | Stoney, Proc. R. Soc. Lond. A 82, 172 (1909). |
| Waveguide sensitivity | Bogaerts & Chrostowski, Laser Photonics Rev. 12 (2018). |
| SPENVIS orbital radiation model | ESA SPENVIS (spenvis.oma.be). Standard tool for mission radiation environment modeling. |
Aether Engine - Coupled multiphysics solver for photonic ICs under extreme environments (hypersonic, cryogenic, radiation, orbital thermal cycling). Open source, Apache 2.0.
Proprietary. Trade secret. See LICENSE for full terms.
Unauthorized copying, reverse engineering, distribution, or use is prohibited under the Defend Trade Secrets Act (18 U.S.C. 1836), the Economic Espionage Act (18 U.S.C. 1831), and the Computer Fraud and Abuse Act (18 U.S.C. 1030).
Licensing inquiries: open a GitHub issue or email contact@aeroza.dev
