Programmable Multi-Wavelength Silicon Photonic Polynomial Evaluator: Eight-Lane Design and Numerical Validation - AELYRION–NERYTH P8

Abstract

AELYRION-NERYTH P8 is a reproducible research and engineering reference for an eight-lane photonic evaluator of real univariate polynomials up to degree seven. The release integrates mathematical derivations, an executable compiler and behavioral simulator, synthetic research data, functional optical and electrical connectivity, calibration procedures, and structured workflows for AI-assisted scientific investigation. The project addresses a central difficulty in analog photonic computation: a polynomial can be mathematically representable while its physical encoding requires excessive dynamic range or becomes sensitive to component losses, gain asymmetry, control uncertainty, and receiver noise. P8 makes those constraints explicit, searches a defined family of implementations, and retains precision-rejected cases alongside successful model examples. PureOne/aelyrion-neryth-p8 · Datasets at Hugging Face This is a pre-fabrication research release. All numerical transfer, current, noise, and transient observations supplied with the project are synthetic. Its verified results concern mathematical identities, software behavior, and conditional device models. Physical accuracy, throughput, energy consumption, and comparative hardware performance require experimental validation. Architecture and mathematical contribution Each lane uses a cascade of seven input-controlled optical splitters to generate eight first-exit power terms. Under the stated affine transfer assumptions, these terms are nonnegative, sum to unity, and form a basis for polynomials of degree at most seven. A deterministic compiler converts a requested polynomial into this basis and encodes its coefficients through separately detected positive and negative branches. The compiler accounts for unequal optical-path transmission, detector response, finite extinction, and asymmetric attainable coefficient ranges. It solves the offset-and-scale feasibility problem for each candidate orientation and examines all 128 degree-seven orientation words. Exact rational comparisons of the supplied binary64 parameters establish the minimum encoding scale within this specified architecture family. The serialized program receives a separate arithmetic residual bound. An upstream optical reference supports electrical reconstruction without introducing the input-dependent denominator that can arise when unequal-loss branches are normalized by their total detected output. Common source-power scaling cancels within the declared reference and gain model; downstream drift and additional receiver errors remain part of the calibration and uncertainty problem. The mathematical documentation presents six guarantees covering polynomial-basis completeness, asymmetric coefficient encoding, finite-family orientation optimization, deterministic perturbation bounds, serialized arithmetic residuals, and conditional noise behavior. These results provide explicit assumptions against which the implementation and future measurements can be checked. Engineering content The functional eight-lane design specifies 56 data Mach-Zehnder interferometers, 64 coefficient interferometers, 128 signal photodiodes, eight reference photodiodes, and 120 slow controls. Eight optical carriers supply independently controlled spatial cores. The connection tables describe 393 optical and 512 electrical connections. The 75-page engineering manual is accompanied by a bench-build guide, functional netlists, control and wavelength tables, receiver models, commissioning templates, and a fabrication handoff checklist. A separate 17-channel branch-readout example demonstrates voltage-to-current conversion and channel validation before summation. These artifacts support process selection, calibration planning, receiver development, and prototype review. Foundry-specific layout, qualified component models, DRC/LVS, packaging, electronics, and measured system acceptance remain open engineering tasks. Reproducible results and retained limitations For seventh-degree smoothstep, orientation optimization reduces the modeled encoding scale from approximately 71.2502 to 19.1406. In the matched nominal scenario, mean modeled noise variance improves by approximately 13.7 times. This is an internal model comparison under the same assumed device resources; the polynomial still fails its requested commissioning tolerance under the more conservative uncertainty allocation. The full-domain shifted Chebyshev T7 example exposes severe conditioning. An eight-segment commissioning program reduces the worst conditional error budget to approximately 0.005546 against an absolute tolerance of 0.01. This improvement uses all eight resident cores for one segmented function, making its resource cost explicit. The release preserves successful programs, precision rejections, receiver guard failures, and the distinction between nominal, commissioning, and idealized scenarios. This allows researchers to investigate where the approach is useful and where its assumptions or resource requirements become restrictive. Dataset and AI-agent support The accompanying Hugging Face Dataset packaging contains 20,114 records across eleven configurations: transfer curves, lane evaluations, nominal benchmarks, commissioning benchmarks, ideal benchmarks, segmentation, thermal transients, control-precision sensitivity, hardware connections, agent tasks, and source documentation. Records carry source provenance, and numerical tables preserve their field meanings and units. The total includes documentary, task, and design records as well as synthetic numerical observations. AI-agent support includes repository instructions, eleven bounded task specifications, reproduction and audit prompts, a source-linked playbook, and a structured reporting contract. These materials support numerical reproduction, compiler inspection, sensitivity studies, calibration planning, and engineering handoff review. They provide portable project guidance for external research and coding agents. Scientific value and intended use The principal contribution is the integration of loss-aware polynomial compilation, reference-normalized signed readout, explicit precision eligibility, and reproducible engineering evidence in one inspectable reference project. It provides a concrete platform for studying analog conditioning, uncertainty propagation, compiler–hardware interactions, calibration requirements, and the transition from a mathematical circuit model to a laboratory implementation. The release is intended for researchers in integrated photonics, optical computing, numerical analysis, analog signal processing, instrumentation, and research software engineering. Its documented prior-art discussion identifies related approaches and frames the proposed combination as a candidate contribution requiring further scrutiny. Status and completeness: The separately recorded fresh numerical reproduction passed 212/212 tests and 11/11 workflow stages, with 56/56 retained artifact-hash checks matching: 100% of those defined checks. Fabricated P8 chips: 0. Measured P8 prototypes: 0. Physical qualification remains open. Engineering source version: 1.2.0. Hugging Face dataset wrapper: 1.2.0-hf1. The source release provides no explicit reuse-license grant; licensing should be resolved before redistribution under a chosen license.