The “wow moment” with physical neural networks is not merely that they run neural computations on unusual chips. It is that, in a literal sense, the physics is the network: light interference, analogue currents, magnetic precession, elastic deformation or bistable fluid–mechanical states directly perform the operations that software normally emulates.
The sober view matters too. Most PNNs are still small-scale demonstrations. Training remains hard; device variation, thermal drift, analogue noise, limited programmability and fabrication yield are recurring bottlenecks. The right conclusion is not “digital is over” — it is that specific parts of neural computation are starting to migrate into physical substrates when the workload matches the physics.
A physical neural network, in the modern literature, is a neural-like computing system in which inference and sometimes learning are carried out by an analogue physical process, rather than being fully simulated on conventional digital hardware. The field now spans five especially useful categories: photonic, memristor-based, spintronic, mechanical, and analogue electronic / mixed-signal systems. Many important systems are hybrids that sit between these buckets.
The taxonomy matters because each class gets its advantage from a different physical law. Photonic systems use propagation, interference, diffraction, wavelength multiplexing and optoelectronic conversion. Memristor systems use programmable conductance states in crossbar arrays. Spintronic systems exploit electron spin, magnetisation dynamics and magnetic tunnel junctions. Mechanical systems use elastic coupling, vibrational modes, bistability and wave propagation in matter.
In photonic PNNs, inputs are encoded in light — usually amplitude, phase, wavelength, time slots or spatial modes. Linear layers are then realised by optical propagation through diffractive elements, Mach–Zehnder interferometer meshes, microring resonators or integrated crossbars. Light propagates with sub-nanosecond latency; interference implements linear algebra directly.
In memristor systems, the core trick is beautifully simple. A weight is stored as device conductance; applying an input voltage produces a current by Ohm’s law; currents on a column sum by Kirchhoff’s law; the array therefore performs vector–matrix multiplication where the weights physically reside. This removes much of the data shuttling that dominates energy use in conventional von Neumann machines.
Spintronic PNNs rely on magnetic degrees of freedom. Magnetic tunnel junctions can store synaptic state non-volatilely; spin-torque nano-oscillators provide nonlinear, oscillatory neuron-like behaviour. Their strength is device-native dynamics, low standby power, and compatibility with memory technologies industry already knows how to manufacture.
Mechanical PNNs are the furthest from the public’s default image of AI hardware and therefore often the most mind-bending. They use physical displacement, stress, vibration, waves or bistability as the state space. Their promise lies in smart materials, adaptive structures, soft robotics and systems where computation and physical response should be the same thing.
Training is where the field gets really interesting. Four broad approaches now coexist: ex-situ/in-silico training; hybrid training using real devices for forward passes; in-situ training that measures gradients directly in hardware; and backpropagation-free or surrogate approaches. The field is converging on an uncomfortable but productive truth: the training algorithm has to respect the substrate’s physics, not merely the other way round.
A photonic chip trained by light itself
Stanford University reported experimentally realised in-situ backpropagation in a three-layer silicon photonic neural network. They measured backpropagated gradients physically by interfering forward- and backward-propagating light. A turning point — moving in-situ optical training from elegant theory to laboratory fact.
The Taichi photonic chiplet
Tsinghua University unveiled Taichi, claiming 160 TOPS/W, “millions-of-neurons capability”, and 91.89% accuracy on the 1,623-category Omniglot dataset. It attacked the usual critique that photonic neural networks look fast but don’t scale.
Single-chip photonic linear + nonlinear layers
A fully integrated coherent photonic deep neural network demonstrated 410 ps latency and reached 92.5% accuracy. Both linear and nonlinear operations lived on the same chip, making the system feel like a real neural processor rather than a linear optical front-end bolted to electronics.
Mechanical network that learns, is damaged, and recovers
University of Michigan introduced the mechanical analogue of in-situ backpropagation. After a bond was pruned, accuracy fell towards 50% — then recovered to ~80% with retraining. One of the clearest demonstrations that “learning matter” is not metaphorical language.
Rack-mounted photonic tensor processor
A photonic tensor processor integrated into a standard 19-inch rack with a high-speed electronic interface to PyTorch, fabricated in imec’s silicon-photonics platform. Demonstrated 98.1% MNIST and 72.0% CIFAR-10 inference — answering not “can you make a beautiful chip?” but “can you place it inside a real computing stack?”
Selected milestones
The table below is intentionally cautious — energy figures often omit I/O, calibration and control overhead, so they should be read as best-paper indicators, not procurement-grade system numbers.
| Modality | Key advantage | Core challenge | Maturity |
|---|---|---|---|
| Photonic | Sub-ns latency, high bandwidth, low energy for linear layers | Nonlinearity, reconfigurability, I/O overhead | Growing |
| Memristor | Dense in-memory compute, removes data shuttling | ADC/DAC overhead, drift, endurance under training | Growing |
| Spintronic | Non-volatile state, low standby power, temporal computing | System-level architecture, benchmark clarity | Early |
| Mechanical | Embodied computation, sensing + compute fused | Task fit, not general throughput | Early |
| Analogue / mixed-signal | Dense, CMOS-compatible, edge inference | Precision vs efficiency trade-off | Most mature |
The photonic camp’s hardest unsolved engineering problem is not “can light multiply matrices?” — that part is already convincing. The hard part is combining nonlinearity, reconfigurability, low loss and scalable cascading in one integrated architecture.
- Photonics: the nonlinearity gapMost on-chip optical neural networks remain constrained by limited input dimensionality, thermal crosstalk, insertion loss, and the energy cost of photoelectric interfaces. Optics is brilliant at fast linear algebra — a full usable neural computer needs much more.
- Memristors: the periphery problemThe core computation is elegant, but simplicity vanishes when counting the periphery. Analog–digital conversion is the primary design trade-off, with energy efficiency and precision pulling in opposite directions.
- Spintronics: the system gapRich in device physics, promising for temporal computing — but the field still needs stronger system-level architectures, better benchmarks and clearer pathways from elegant oscillators to full trainable networks.
- Mechanical: task fit over throughputMechanical PNNs will not win by imitating transformer accelerators poorly. They win by solving problems where physical response, memory and control should already be co-located.
- Universal: training robustnessDevice mismatch, fabrication variation, thermal drift, phase error, misalignment and non-ideal peripheral circuits can invalidate a model after deployment. Robustness is not a side issue — it is the issue.
Commercially, the near-term picture is uneven but no longer hypothetical. The clearest momentum is in photonics for AI infrastructure, especially interconnect, chiplet fabrics and specialised linear algebra. Among companies, Lightmatter, Lightelligence and Celestial AI are all positioning around light-based AI hardware, while Mythic already sells an analogue matrix processor for edge inference.
The most plausible first use-cases are narrower niches: datacentre optical interconnect; edge inference where memory movement dominates power; always-on sensing and temporal processing; RF and communications tasks; and robotics controllers where the controller is physically embodied in the structure.
The ethics and safety questions are more concrete than the field’s hype sometimes suggests. PNNs could improve privacy if more inference moves onto edge devices. But hardware drift can silently degrade performance after deployment, and where computation is coupled directly to actuation, a “model bug” can become a physical failure mode.
Robust, transferable training pipelines
Methods that survive fabrication variation, thermal drift and device mismatch will matter more than another isolated headline demo.
Nonlinearity without killing the energy story
In photonics especially, scalable nonlinear and reconfigurable layers are the hinge between “beautiful linear accelerator” and “useful neural computer”.
Packaging and software co-design
Integration with PyTorch stacks, FPGAs, chiplets and packaging may move the field faster than purist substrate advances alone.
Honest end-to-end metrics and standards
The field needs benchmark suites that count I/O, converters, calibration, drift management and retraining — not only idealised core operations.
Embodied PNNs that do not merely imitate GPUs
Mechanical and multistable systems are most exciting when they fuse sensing, memory, computation and action into adaptive materials and robots.
Primary-source reading list
- All-optical machine learning using diffractive deep neural networks (2018) — Science
- Wave physics as an analog recurrent neural network (2019)
- Fully hardware-implemented memristor convolutional neural network (2020)
- Deep physical neural networks trained with backpropagation (2022) — Nature
- Experimentally realised in situ backpropagation for deep learning in photonic neural networks (2023)
- All-analog photoelectronic chip for high-speed vision tasks (2023) — Nature
- Single-chip photonic DNN with forward-only training, 410 ps latency (2024) — Nature Photonics
- Training all-mechanical neural networks through in situ backpropagation (2024) — Nature Communications
- Training of physical neural networks — 2025 Nature review — Nature
- Physical neural networks using sharpness-aware training (2026)
Open questions remain. The field still lacks apples-to-apples comparisons across modalities; many reported tasks are small or highly structured; and some “PNN” systems are really hybrids whose efficiency depends on opaque off-chip electronics. The right near-term expectation is not wholesale replacement of digital AI hardware, but a gradual capture of the neural-network operations that specific physical substrates perform unusually well. That is already enough to make PNNs one of the most intellectually alive corners of AI hardware today.
