phys-MCP: A Control Plane for Heterogeneous Physical Neural Networks
Integrating biological, photonic, and molecular neural networks into edge computing just got a universal interface.
Physical neural networks (PNNs) embed computation directly into material dynamics — ranging from photonic circuits and memristors to molecular and even biological wetware. They promise ultra-low-power, real-time computation at the extreme edge, where sensing and actuation meet the physical world. But until now, each PNN substrate has required custom software interfaces for timing, observability, lifecycle management, and integration with cloud or edge workflows. This fragmentation has kept PNNs from being practically deployed alongside conventional AI stacks.
phys-MCP solves this by acting as a universal control plane. It models each substrate's capabilities — latency, resetability, plasticity, I/O modality — and exposes them through standard APIs. The architecture supports digital twin bindings for simulation and telemetry-driven fault recovery. The prototype validates against three backend classes: a conventional HTTP-driven stack, and more critically, a Cortical Labs wetware adapter that exposes living biological neurons as programmable compute resources. Test results confirm that phys-MCP enables middleware-agnostic integration, improves runtime-aware backend matching over simpler baselines, and recovers gracefully from representative faults — all with minimal local control-path overhead.
- Supports heterogeneous PNNs including molecular, chemical, biological, photonic, memristive, and mechanical substrates.
- Prototype includes a Cortical Labs adapter that exposes wetware (biological neural networks) through the same control model.
- Demonstrates telemetry-aware recovery under representative faults and runtime-aware backend matching with small overhead.
Why It Matters
phys-MCP paves the way for practical edge AI using exotic materials, from biological neurons to photonic circuits.