Physical Neural Networks: Analog AI Systems

• Physical Neural Networks are neural-inspired computing platforms that leverage physical substrates like optical circuits, memristors, and mechanical systems for ultrafast, energy-efficient processing.
• PNNs implement key neural operations through analog linear mixing and device-specific nonlinearity such as threshold switching and bistability, emulating activations and weight storage.
• Emerging training protocols, including hybrid physics-aware methods and backpropagation-free local learning, address hardware nonidealities while supporting scalability and robustness.

A Physical Neural Network (PNN) is a neural-like computational system where the core information processing—corresponding to neuron activations and synaptic weights—occurs via analog transformations in physical substrates such as optical, electronic, mechanical, or chemical media, rather than in virtualized digital arithmetic. PNNs leverage the native dynamics, nonlinearity, and energy efficiency of physical phenomena, enabling in-memory, parallel, and ultrafast computation. Research over the past decade has established PNNs as a major direction for energy-efficient AI hardware, with diverse implementations, unique training algorithms, and domain-specific architectures (Momeni et al., 2024, Momeni et al., 2023, Langer, 2024, Yu et al., 2024).

1. Physical Realizations and Architectures

PNNs are categorized by the physical infrastructure underlying neuron and synapse analogues, the manner in which programmability and nonlinearity are achieved, and the structural mapping to conventional neural-network layers:

• Isomorphic PNNs: Systems such as memristor crossbars or programmable photonic meshes, where the physical network is in direct, mathematical correspondence with the operations of classical neural nets (e.g., matrix-vector multiplication). Here, weights are mapped to, e.g., conductances or phase shifts, and activations to voltages or light intensities (Momeni et al., 2024, Wang et al., 10 Feb 2026).
• Broken-isomorphism PNNs: Systems utilizing complex or disordered analog dynamics (e.g., optical scattering media, mechanical wave lattices) where the input-output mapping realizes a nonlinear transformation not easily decomposed into digital layers (Momeni et al., 2023, Stenning et al., 2022).

Key physical domains include:

• Optical/Photonic PNNs: Integrated Mach-Zehnder interferometers, microring resonators, and monolithic lithium niobate optical processors implement linear transforms and nonlinearities via interference, phase modulation, and optical nonlinear effects (Liu et al., 28 Jul 2025, Xu et al., 2024).
• Memory-resistor (Memristor/PCM) Arrays: Crossbar topologies permit direct matrix computations with synaptic weights encoded in analog conductance states (Wang et al., 10 Feb 2026, Sakemi et al., 2024).
• Mechanical/Metamaterial Networks: Mechanical bistability or wave propagation in structured media yield architectures supporting computation, memory, and actuation (Ben-Haim et al., 2024, Li et al., 10 Mar 2025).
• Polymeric and Chemical Networks: Nanostructured polyaniline networks or ionic/electrochemical devices enable self-assembled connectivity and redox-based programming (Langer, 2024).

2. Signal Propagation and Nonlinearity Mechanisms

PNNs implement the signature neural operations through:

• Analog Linear Mixing: Physical electrical or optical circuits distribute input signals according to a set of weights, embodied as conductances, refractive indices, or spring constants. For instance, in a nanowire network, the effective conductance $G_{ij}$ between input $i$ and output $j$ implements weighted summation (Langer, 2024, Momeni et al., 2024).
• Element-wise Nonlinearity: Physical PNNs employ device-specific mechanisms:
  • Threshold Switching: Redox transitions in conductive polymers (PANI), memristive filaments, or optoelectronic saturable absorption (Langer, 2024, Xu et al., 2024).
  • Wave Interference and Saturation: Nonlinear response in photonics and mechanical lattices via material thresholds or band gaps (Liu et al., 28 Jul 2025, Li et al., 10 Mar 2025).
  • Bistability: Double-well mechanical potentials in elastic chambers enable binary state storage and all-or-nothing transformations (Ben-Haim et al., 2024).

Collectively, these analogues furnish PNNs with "activation function" behavior, e.g., abrupt changes in conductance mimicking neuronal firing.

3. Learning Protocols and Training Algorithms

Training PNNs departs from classical digital deep learning due to physical constraints and lack of universal differentiability. Diverse strategies have been developed:

• Hybrid Physics-Aware Training (PAT): Forward passes are run on the real physical device, while gradients are estimated using a digital model (or "digital twin") for the backward pass. This approach permits backpropagation despite hardware nonidealities but is limited by the accuracy of the model and requires access to device-specific parameterizations (Wright et al., 2021, Momeni et al., 2024).
• Sharpness-Aware Training (SAT): A two-step optimization (adversarial step plus descent) seeks broad ("flat") minima in the loss landscape, improving robustness to fabrication, drift, and perturbations, and supporting parameter transferability across PNN instances (Xu et al., 2024).
• Backpropagation-Free Local Learning: Techniques such as Model-Free Forward-Forward (MF-FF) training employ layer-local, contrastive learning rules that use pairs of forward passes (on true and "wrong" labelings) to update the trainable weights, completely bypassing end-to-end gradient flow (Momeni et al., 2023). Information Bottleneck-based training applies layerwise mutual information objectives to optimize for task relevance and noise robustness (Wang et al., 10 Feb 2026).
• Physical Self-Learning and Contrastive Protocols: Some PNNs employ intrinsic physical adaptation, with weight updates driven by built-in feedback mechanisms such as local redox changes, Hebbian current reinforcement, or mechanical plasticity, requiring little or no digital supervision (Yu et al., 2024, Langer, 2024, Ben-Haim et al., 2024, Li et al., 10 Mar 2025).
• Biologically Plausible Learning: Approaches like Direct Feedback Alignment (DFA) and Equilibrium Propagation utilize random feedback or energy-based updates compatible with local physical operations (Sunada et al., 26 Feb 2025, Yu et al., 2024).

The table below summarizes key training paradigms and their requirements:

Method	Key Idea	Gradient Source
Physics-Aware Training	Digital model backward	Digital twin
Sharpness-Aware Training	Flat minima optimization	Device or model, 2-pass
Forward-Forward / Local	Two forward passes	Local, contrast
Self-learning	Physical adaptation	Substrate-intrinsic

4. Device Programming, Endurance, and Parallelism

PNN programmability is realized via physical parameter manipulation:

• Electrical Programming: Pulse-based potentiation or depression of memristor/PCM crossbar elements, phase-shifter bias tuning in photonic circuits, or current-induced redox in polymeric networks (Xu et al., 2024, Langer, 2024).
• Chemical/Growth Programming: Directing nanowire growth by electrochemical bias to preconfigure network topology (Langer, 2024).
• Mechanical Programming: Changing conductances or stiffness by mechanical actuation, plastic deformation, or bistable switching (Ben-Haim et al., 2024).

Programming endurance varies by substrate, with cycle limits depending on fatigue (mechanics), redox reversibility (chemistry), or drift (electronic/photonic). Typical endurance for laboratory PNNs ranges from $10^3$ (for PANI) to $10^{12}$ (for photonic phase shifters).

PNNs leverage massive parallelism: entire inputs are broadcast and propagate simultaneously through passive networks; outputs emerge as analog signals requiring no sequential readout or multiplexing.

5. Performance Metrics and Energy Efficiency

Experimental PNNs report:

• Inference Latency: Picosecond to millisecond timescales depending on substrate (optical < nanosecond, mechanical > ms) (Liu et al., 28 Jul 2025, Langer, 2024, Ben-Haim et al., 2024).
• Energy per MAC: Photonic and resistive PNNs achieve 0.5–50 fJ/MAC, well below digital CMOS (10 pJ–1 nJ/MAC) (Liu et al., 28 Jul 2025, Sakemi et al., 2024, Xu et al., 2024).
• Robustness: Local learning and SAT confer resilience to thermal, noise, and drift perturbations, with transfer capability across device instances (Xu et al., 2024, Wang et al., 10 Feb 2026).
• Scalability: Demonstrated networks range from <10 (chemomechanical) up to $10^4$ parameters (photonic mesh, memristive crossbar), with emerging schemes (e.g., ReLaX-Net time-multiplexed reuse) targeting further scaling (Tsuchiyama et al., 28 Oct 2025).
• Accuracy: On standard tasks (MNIST, Fashion-MNIST, AG News), photonic and memristor PNNs approach or match classical DNN performance: $>$96% on MNIST, $>$86% on Fashion-MNIST (Liu et al., 28 Jul 2025, Wang et al., 10 Feb 2026).

6. Limitations, Challenges, and Future Directions

PNNs face challenges in mass manufacturability, parameter resolution, drift, device variability, and hybrid integration with digital logic:

• Reproducibility and Scaling: Batch-to-batch variability in self-assembled/chemical systems; need for sub-wavelength lithography or hybrid CMOS/selector integration for high-density arrays (Langer, 2024, Xu et al., 2024).
• Drift and Nonideality Compensation: Time-varying behavior from temperature, aging, or device mismatch requires robust training (e.g., SAT, PIB) and hardware-aware retraining (Xu et al., 2024, Wang et al., 10 Feb 2026).
• Resolution and Noise Floor: Physical substrates are limited by finite device resolution (e.g., few bits of analog precision) and noise, motivating architectures that exploit robustness rather than rely on high-precision computation (Xu et al., 2024, Liu et al., 28 Jul 2025).
• Programmable Depth and Parameter Bottlenecks: Parameter-efficient architectures such as ReLaX-Net employ time-multiplexed layer reuse to maximize effective model depth relative to slow weight reconfiguration timescales (Tsuchiyama et al., 28 Oct 2025).

Future research targets include scalable local learning rules for deep/broad PNNs, hybrid optoelectronic architectures, rapid programming techniques, self-adaptive device arrays, and integration with edge computing and in-sensor inference (Momeni et al., 2024, Tsuchiyama et al., 28 Oct 2025, Liu et al., 28 Jul 2025).

7. Interpretability and Physics-Constrained Computation

PNNs are also leveraged for physics-informed tasks and interpretable modeling:

• Physical Law Discovery: Parsimonious neural networks employ evolutionary selection for compact architectures, enabling recovery of explicit physical laws (e.g., symplectic integrators, melting-temperature scaling) directly from data (Desai et al., 2020).
• Physics-Constrained Activations and Architectures: Networks can be constructed to obey differential constraints via tailored activations, kernel methods, or by explicit encoding of physical priors (e.g., symplecticity in Poisson neural networks) (Ranftl, 2022, Jin et al., 2020).

These approaches bridge data-driven learning with physical interpretability, providing a framework for model induction in scientific domains.

---

References:

(Langer, 2024): Programmable polyaniline nano neural network. A simple physical model (Momeni et al., 2023): Backpropagation-free Training of Deep Physical Neural Networks (Ben-Haim et al., 2024): Multistable Physical Neural Networks (Wang et al., 10 Feb 2026): Training deep physical neural networks with local physical information bottleneck (Stenning et al., 2022): Neuromorphic Overparameterisation and Few-Shot Learning in Multilayer Physical Neural Networks (Sakemi et al., 2024): Harnessing Nonidealities in Analog In-Memory Computing Circuits: A Physical Modeling Approach for Neuromorphic Systems (Tsuchiyama et al., 28 Oct 2025): ReLaX-Net: Reusing Layers for Parameter-Efficient Physical Neural Networks (Liu et al., 28 Jul 2025): Monolithically Integrated Optical Convolutional Processors on Thin Film Lithium Niobate (Li et al., 10 Mar 2025): Topological mechanical neural networks as classifiers through in situ backpropagation learning (Xu et al., 2024): Perfecting Imperfect Physical Neural Networks with Transferable Robustness using Sharpness-Aware Training (Momeni et al., 2024): Training of Physical Neural Networks (Xu et al., 2024): Control-free and efficient integrated photonic neural networks via hardware-aware training and pruning (Ranftl, 2022): A connection between probability, physics and neural networks (Sunada et al., 26 Feb 2025): Blending Optimal Control and Biologically Plausible Learning for Noise-Robust Physical Neural Networks (Jin et al., 2020): Learning Poisson systems and trajectories of autonomous systems via Poisson neural networks (Wright et al., 2021): Deep physical neural networks enabled by a backpropagation algorithm for arbitrary physical systems (Yu et al., 2024): Physical Neural Networks with Self-Learning Capabilities (Desai et al., 2020): Parsimonious neural networks learn interpretable physical laws