cs.NEApr 10, 2026

Beyond Silicon: Materials, Mechanisms, and Methods for Physical Neural Computing

Authors: Stefan Fischer, Nihat Ay, Olaf Landsiedel, Esfandiar Mohammadi, Sebastian Otte, Bernd-Christian Renner, Nele Rußwinkel

Organizations: Institute for Telematics, University of Lübeck, Germany · Institute for Data Science Foundations, TU Hamburg, Germany · Institute for Networked Cyber-Physical Systems, TU Hamburg, Germany · United Nations University Hub on Engineering to Face Climate Change at TU Hamburg, United Nations University Institute for Water, Environment and Health (UNU-INWEH), Hamburg, Germany · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Physical implementations of neural computation now extend far beyond silicon hardware, encompassing substrates such as memristive devices, photonic circuits, mechanical metamaterials, microfluidic networks, chemical reaction systems, and living neural tissue. By exploiting intrinsic physical processes such as charge transport, wave interference, elastic deformation, mass transport, and biochemical regulation, these substrates can realize neural inference and adaptation directly in matter. As silicon GPU-centered AI faces growing energy and data-movement constraints, physical neural computation is becoming increasingly relevant as a complementary path beyond conventional digital accelerators. This trend is driven in particular by pervasive intelligence, i.e., the deployment of on-device and edge AI across large numbers of resource-constrained systems. In such settings, co-locating computation with sensing and memory can reduce data shuttling and improve efficiency. Meanwhile, physical neural approaches have emerged across disparate disciplines, yet progress remains fragmented, with limited shared terminology and few principled ways to compare platforms. This survey unifies the field by mapping neural primitives to substrate-specific mechanisms, analyzing architectural and training paradigms, and identifying key engineering constraints including scalability, precision, programmability, and I/O interfacing overhead. To enable cross-domain comparison, we introduce a first-order benchmarking scheme based on standardized static and dynamic tasks and physically interpretable performance dimensions. We show that no single substrate dominates across the considered dimensions; instead, physical neural systems occupy complementary operating regimes, enabling applications ranging from ultrafast signal processing and in-memory inference to embodied control and in-sample biochemical decision making.

Figures & tables

Figure 1: Landscape of physical substrates for computation and learning. The horizontal axis organizes approaches by the dominant coupling carrier (charge and electronics, waves and fields, and mass and matter transport), while the vertical axis groups them by substrate class and physical scale (molecular and wet, engineered solid-state, and macro and embodied systems). Each box summarizes a representative paradigm and indicates the corresponding chapter(s). Across this landscape, material physics determines which neural primitives are natively available in a substrate, whereas architecture determines how these primitives are arranged into scalable computational organizations.
Figure 2: Mechanisms of nucleic acid strand displacement. (A) Three-way branch migration: Two identical strands compete for binding to a common complementary template. (B) Stepwise kinetics: The migration proceeds as an unbiased random walk via the dissociation and reformation of single base pairs at the branch point. (C) Toehold-mediated strand displacement (TMSD): An invader strand binds to a single-stranded overhang ("toehold") to efficiently initiate branch migration and displace the incumbent strand. (D) Secondary structure invasion: An invader opens a hairpin loop via displacement, activating a previously sequestered sequence domain (blue). Reprinted from [ 125 ] , licensed under CC BY-NC 4.0.
Figure 3: Architecture of a Chemical Reservoir Computer based on the Formose Reaction. (a) The chemical reaction network functions as a physical reservoir (a fixed, non-linear hidden layer). It projects low-dimensional inputs (reactants) into a high-dimensional chemical state space. This intrinsic non-linearity enables the linear separability of complex computational tasks, such as classification and time-series forecasting. (b–f) Signal Flow and Readout Training. A time-dependent input signal u(t)u(t) modulates the inflow rates of reagents into a continuous stirred-tank reactor (CSTR). The resulting non-linear chemical dynamics (the state vector xx ) are monitored in real-time via mass spectrometry. Consistent with the reservoir computing paradigm, training is restricted to the readout layer: a linear weight matrix WW is optimized via regression to map the high-dimensional chemical state to the target output function ( y≈Wxy\approx Wx ). Reprinted from [ 7 ] , licensed under CC BY 4.0.
Figure 4: Schematic of the closed-loop “Pong” experiment. Human cortical neurons cultured on a high-density multielectrode array (HD-MEA) are bidirectionally coupled to a simulated game environment: game states are encoded as electrical stimulation patterns, neural responses are recorded and decoded as motor output, and structured feedback is returned to the culture. The setup illustrates how synthetic biological intelligence can realize real-time learning through closed-loop interaction aimed at reducing environmental unpredictability. Reprinted from [ 59 ] , licensed under CC BY 4.0.
Figure 5: Capacitive in-sensor tactile computing architecture. (a) Schematic of the flexible capacitive sensor array integrated with electrical switches. (b) Circuit diagram demonstrating in-situ multiplication and accumulation (MAC) operations within a sensor subregion. (c) Equivalent tactile neural network mapping input sensor capacitance vectors CC to voltage weights VV . (d) Examples of low-level sensory tasks, including noise reduction, edge detection, and sharpness extraction, computed directly within the sensor layer. The figure illustrates how early neural processing can be embedded directly into the sensing substrate, thereby reducing data transfer to external processing units. Reprinted from [ 18 ] , licensed under CC BY 4.0.
Figure 6: Working principle of a 4-node self-learning magnetic Hopfield neural network (HNN). (A) Schematic of the experimental device structure. (B) Voltage pulses applied to the top Au layer encode the input or boundary conditions and generate current distributions that configure the spin orientation in the Py layer via the Oersted field (C). The resulting magnetic texture relaxation defines the effective conductance matrix, mimicking self-adjusting synaptic weights. (D) The anisotropic magnetoresistance (AMR) effect is used to read out the evolving network state through the angle-dependent relation between current and magnetization. The figure illustrates how associative memory and gradient-descent-like learning emerge directly from ferromagnetic energy minimization. Reprinted from [ 89 ] , licensed under CC-BY-NC-ND 4.0.
Figure 7: Diffractive deep neural networks (D 2 NNs). (A) A D 2 NN uses multiple trainable diffractive layers with complex-valued transmission or reflection coefficients to implement an optical function between input and output planes. Once the layer parameters have been optimized and the structure has been fabricated, inference is performed passively through wave propagation at the speed of light. (B) Example of a D 2 NN for handwritten digit classification. (C) D 2 NN functioning as an imaging lens. (D) Comparison with a conventional neural network, highlighting coherent complex-valued inputs and diffraction-based layer interconnections. The figure illustrates the key trade-off of this photonic regime: extremely fast and energy-efficient inference at the cost of limited post-fabrication programmability. Reprinted from [ 75 ] with permission from the American Association for the Advancement of Science.
Figure 8: Mechanical neural network (MNN) fabricated for the experimental study of Lee et al. [ 70 ] . (A) Full-view photograph of the lattice-based mechanical network. (B) Close-up of the beam structure whose local stiffness values act as trainable mechanical weights. The figure illustrates the physical realization of a mechanical perceptron-like architecture in which desired behaviors, such as shape morphing, are achieved by optimizing the geometry and stiffness distribution of the lattice. Reprinted from [ 70 ] with permission from the American Association for the Advancement of Science.
Figure 9: Physical reservoir computing using a soft silicone arm. (a) Information-processing loop: motor commands actuate the arm, while embedded bend sensors capture the resulting body dynamics as reservoir states. (b) Experimental prototype of the soft arm. (c) Conceptual comparison showing how the physical soft body replaces the abstract reservoir layer of a conventional Echo State Network. The figure illustrates the central idea of embodied reservoir computing: computation emerges from the intrinsic transient dynamics of the compliant physical system, while only the readout is trained. Reprinted from [ 87 ] , licensed unter CC-BY 4.0.
Figure 10: Artificial Hodgkin-Huxley-like neuron based on 2D ionic memristors. (A) Circuit representation of the Hodgkin-Huxley model showing the ionic branches and corresponding Nernst potentials. (B) Brownian-dynamics simulation of the ionic prototype, consisting of two slits connected to reservoir pairs with different ionic concentrations. (C) Spontaneous voltage spikes generated by the prototype, with the right panel illustrating the quiescent, initiation, and discharge phases of a single spike. The figure exemplifies how nanofluidic transport asymmetry and ionic coupling can reproduce biologically familiar excitability in a fully physical device. Reprinted from [ 112 ] with permission from the American Association for the Advancement of Science.
Protocol elementStatic classification taskDynamic prediction/control taskReporting requirement
Input dimensionalityReference input: MNIST-style 28×2828\times 28 images where feasible; compressed variants with d∈{16,64,256}d\in\{16,64,256\} for substrates with limited I/O.Low-dimensional time series with d∈[1,16]d\in[1,16] input channels; sequence length T=100T=100 – 10001000 samples or substrate-specific physical time horizon.Report the physical encoding of each input dimension, compression method if used, and whether dimensions are applied in parallel or sequentially.
Noise levelsEvaluate clean inputs and additive input/readout noise at representative SNR levels, e.g., 40 dB, 20 dB, and 10 dB where experimentally meaningful.Evaluate process or readout perturbations at matched SNR levels; for closed-loop tasks additionally report disturbance amplitude and duration.State whether noise is injected digitally, physically at the input, internally by the substrate, or only at readout; report mean and variance over repeated trials.
Training regimeReport whether parameters are obtained by ex-situ digital training, device-aware training, in-situ adaptation, reservoir readout training, or closed box optimization.Same categories; additionally report whether temporal memory is intrinsic to the substrate or provided by external delay/state variables.Specify train/test split, number of training examples or episodes, number of physical evaluations, and whether the same physical instance is reused.
Performance metricClassification accuracy, confusion matrix, and robustness degradation under noise/compression.Prediction error (e.g., NRMSE) for time-series tasks; task reward, tracking error, or stability margin for control tasks.Report mean, standard deviation, number of runs, and whether variability is dominated by device noise, fabrication spread, biological variability, or environmental drift.
Latency and throughputTime from input presentation to stable or readable output, including settling or relaxation time.Sampling interval, memory horizon, closed-loop reaction time, and maximum stable control/update rate.Separate intrinsic substrate time constants from encoding, actuation, readout, and digital post-processing latency.
Energy accountingEnergy per inference or per classified sample.Energy per prediction step, control step, or physical episode.State the accounting boundary: core substrate only, substrate plus actuation/readout, or full system including ADC/DAC, pumps, lasers, cryogenics, incubation, or life support.
Table I: Minimal reference protocol for cross-substrate benchmarking of physical neural systems. The proposed values are intended as reporting defaults rather than rigid standards; substrate-specific deviations should be stated explicitly.
Figure 12
Figure 13
Figure 14
Figure 15
Figure 16
Figure 17
Figure 18
More in this topic

Neuromorphic Computing

Neuromorphic computing uses brain-inspired, event-driven algorithms and hardware to process information with low power and latency.

See all 186 papers

Explore similar work

CardsList
May 5, 2026

CP2^2N2^2: A Control Plane for Heterogeneous Physical Neural Networks

Physical neural networks (PNNs) exploit diverse physical substrates---spanning molecular, wetware, and solid-state devices---to perform embodied neural computation close to physical processes. However, their operational heterogeneity in signal modalities, timing regimes, reset requirements, calibration lifecycles, and observability prevents their integration as interchangeable execution endpoints in edge, fog, and cloud workflows. We present CP2^2N2^2, a substrate-aware control-plane architecture that bridges software-level orchestration and physical neural computing. CP2^2N2^2 introduces a three-plane model that decouples high-level orchestration, twin-state synchronisation, and substrate-specific execution under versioned Physical Neural Resource Contracts. It extends Model Context Protocol (MCP)-style discovery and invocation with explicit lifecycle management, atomic leases, telemetry freshness verification, and fail-closed safety boundaries. We evaluate a reference prototype across contrasting behavioural twins, a same-host multi-process deployment under fault injection, and the Cortical Labs SDK Simulator. In a 945-request concurrency campaign, CP2^2N2^2 enforces typed contention and fail-closed recovery with sub-millisecond local orchestration overhead. Furthermore, in a 160-decision Agent-to-PNN campaign with a hosted LLM planner, independent enforcement ensured that 100% of executed actions remained strictly safe, verified, and reconciled despite planner-level inconsistencies. CP2^2N2^2 thus provides an enforceable systems abstraction for integrating scarce, stateful physical AI substrates into agentic computing environments.

Stefan Fischer, Maliheh Hariri, Sebastian Otte

University of Luebeck, Institute of Telematics · University of Luebeck, Institute of Robotics and Cognitive Systems

Heterogeneous ComputingPhysical Neural Networks
AskPDF
Jul 30, 2026

Nanoparticle Networks for Neuromorphic Computing

Physical computing leverages complex dynamical systems for energy-efficient data processing. In this work, we present a neuromorphic architecture based on metallic nanoparticles interconnected by molecular junctions on a SiO2\text{SiO}_2/Si substrate. We demonstrate that surrounding static control electrodes transform this nanoparticle network from a passive reservoir into a tunable nonlinear dynamical system. By analyzing how these electrodes route simple one-dimensional voltage inputs into multidimensional signal responses, we establish three core design rules to maximize computational performance. First, operating near the system's cutoff frequency achieves an optimal balance between nonlinear charge tunneling and linear capacitive memory. Second, tuning the underlying SiO2\text{SiO}_2 thickness sets the electrostatic screening length and dictates the memory type. Thick oxide layers reduce the screening length, causing networks larger than this length to transition into a persistent, non-volatile-like regime. Conversely, networks smaller than the screening length exhibit only fading memory. Third, introducing structural disorder via heterogeneous molecular junctions overcomes inherent limits on expressivity. While a network's computational expressivity scales with its physical size, it is ultimately capped by the screening length. Breaking internal spatial symmetries with localized disorder bypasses this saturation, allowing control voltages to independently manipulate specific signal amplitudes and phases, universally maximizing performance for dynamic neuromorphic applications.

Jonas Mensing, Wilfred G. van der Wiel, Andreas Heuer

Theory of Complex Systems Group, Institute of Physical Chemistry, University of Münster, Münster, Germany · NanoElectronics Group, MESA+ Institute for Nanotechnology, Center for Brain-Inspired Nano Systems (BRAINS), University of Twente, Enschede, Netherlands · Institute of Physics, University of Münster, Münster, Germany

Reservoir ComputingNeuromorphic Computing
AskPDF
Mar 19, 2026

Brain-inspired AI for Edge Intelligence: a systematic review

While Spiking Neural Networks (SNNs) promise to circumvent the severe Size, Weight, and Power (SWaP) constraints of edge intelligence, the field currently faces a "Deployment Paradox" where theoretical energy gains are frequently negated by the inefficiencies of mapping asynchronous, event-driven dynamics onto traditional von Neumann substrates. Transcending the reductionism of algorithm-only reviews, this survey adopts a rigorous system-level hardware-software co-design perspective to examine the 2020-2025 trajectory, specifically targeting the "last mile" technologies - from quantization methodologies to hybrid architectures - that translate biological plausibility into silicon reality. We critically dissect the interplay between training complexity (the dichotomy of direct learning vs. conversion), the "memory wall" bottlenecking stateful neuronal updates, and the critical software gap in neuromorphic compilation toolchains. Finally, we envision a roadmap to reconcile the fundamental "Sync-Async Mismatch," proposing the development of a standardized Neuromorphic OS as the foundational layer for realizing a ubiquitous, energy-autonomous Green Cognitive Substrate.

Yingchao Cheng, Meijia Wang, Zhifeng Hao +1

Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen, 518107, China · Cloud Computing and Distributed Systems (CLOUDS) Lab, School of Computing and Information Systems, University of Melbourne, Melbourne, VIC 3010, Australia · College of Mathematics and Computer Science, Shantou University, Shantou, 515063, China

Neuromorphic ComputingSpiking Neural NetworksEnergy-Efficient MLEdge ML
AskPDF