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

1. Problem Statement

The dominance of silicon-based Artificial Neural Networks (ANNs) is facing fundamental constraints. As Moore's Law and Dennard Scaling approach asymptotic limits, the "Von Neumann bottleneck"—the energy and latency cost of shuttling data between separate processing and memory units—has become a primary barrier. Furthermore, the rise of pervasive intelligence (on-device and edge AI) demands energy-efficient solutions that conventional digital accelerators struggle to provide.

While neuromorphic engineering has attempted to bridge this gap using silicon (e.g., spiking neural networks, analog circuits), these approaches remain constrained by the physics of electron transport in rigid lattices. They often simulate neural dynamics rather than embodying them. The paper argues that alternative physical substrates (biological, chemical, mechanical, photonic, etc.) offer intrinsic capabilities—such as thermodynamic relaxation, wave interference, and mass transport—that can perform neural inference and adaptation directly in matter, potentially bypassing the transduction tax and data-movement costs inherent in digital systems. However, the field is currently fragmented, lacking shared terminology, principled comparison methods, and a unified understanding of how different substrates map to neural primitives.

2. Methodology

This paper is a comprehensive survey and unification of the field of Physical Neural Networks (PNNs). The authors define PNNs as systems where training and inference exploit the intrinsic dynamics of a material substrate rather than relying solely on numerical operations on digital processors.

The methodology involves:

• Taxonomy Construction: Classifying diverse substrates not merely by material (wet vs. dry) but by computational characteristics: signal encoding (data type), plasticity implementation (memory architecture), and temporal dynamics (execution model).
• Mechanism Mapping: Systematically mapping core neural primitives (weighted summation, nonlinearity, memory, adaptation) to the governing physical laws of specific substrates (e.g., Fick's laws for diffusion, Kirchhoff's laws for current, reaction-diffusion equations for chemistry).
• Training Paradigm Analysis: Categorizing learning strategies into three broad classes: ex-situ digital-twin optimization, hybrid digital-physical training (e.g., reservoir computing), and in-materio learning (physical gradient approximation).
• Comparative Benchmarking: Introducing a first-order framework for cross-substrate comparison based on standardized static and dynamic tasks and physically interpretable performance dimensions (speed, energy, plasticity, and I/O overhead).

3. Key Contributions

A. Unified Framework and Taxonomy

The paper establishes a coherent research direction by mapping neural primitives to physical mechanisms across a diverse landscape of substrates, organized into four conceptual parts:

1. Foundations: Defining PNNs relative to neuromorphic engineering and digital ML, and mapping neural operations to physical laws (transport, interaction, relaxation).
2. Wetware and Molecular Intelligence: Surveying DNA strand displacement, chemical reaction-diffusion systems, and synthetic biological intelligence (including organoids and Physarum).
3. Physical Neural Hardware: Reviewing engineered non-biological platforms, including memristive/ferroelectric in-memory computing, spintronic/superconducting systems, photonic circuits, mechanical metamaterials, microfluidic networks, and iontronics.
4. Benchmarking and Outlook: Proposing a comparative framework to evaluate these disparate technologies.

B. Mapping Neural Primitives to Physics

The authors detail how specific physical phenomena realize neural functions:

• Weighted Summation: Achieved via linear superposition (e.g., Kirchhoff's current law in crossbars, diffusion in fluids, wave interference in photonics).
• Nonlinearity: Realized through material instabilities (buckling in mechanics), saturation kinetics (enzymatic reactions), or thresholding (Josephson junctions, memristive switching).
• Optimization/Learning: Framed as thermodynamic relaxation toward energy minima (e.g., DNA binding, spin glass relaxation) or physical gradient descent (e.g., equilibrium propagation).

C. Training and Design Paradigms

The survey categorizes learning strategies based on substrate constraints:

• Ex-situ (Digital Twin): Training a differentiable model in software and transferring weights to hardware. Effective for substrates with high-fidelity models but suffers from the "reality gap."
• Hybrid (Reservoir Computing): Exploiting the fixed, rich dynamics of the physical substrate as a non-linear projection space, training only the readout layer. This is substrate-agnostic and robust to internal variability.
• In-materio (Physical Learning): Updating weights directly within the physical system (e.g., via equilibrium propagation or local Hebbian rules). This closes the reality gap but faces engineering challenges in observability and update speed.

D. Benchmarking Framework

Recognizing that a single scoreboard is insufficient, the authors propose a first-order benchmarking scheme involving:

• Standardized Tasks: Static classification (e.g., MNIST) and dynamic prediction/control (e.g., Mackey-Glass, NARMA).
• Physically Interpretable Dimensions: Reporting speed (Hz), energy per operation (J/op), plasticity mechanisms, and crucially, I/O and interface overhead.
• Reporting Protocol: A minimal reference protocol (Table I) mandating transparency regarding input encoding, noise levels, training regimes, and energy accounting boundaries (e.g., core physics vs. full system including pumps/lasers).

4. Results and Comparative Landscape

The survey synthesizes findings from over a decade of research, revealing that no single substrate dominates across all dimensions. Instead, physical neural systems occupy complementary operating regimes:

• Molecular/DNA: Offers extreme energy efficiency and massive parallelism (thermodynamic limits) but is limited by slow diffusion kinetics (minutes to hours) and readout challenges. Best for in-sample diagnostics.
• Chemical/Reaction-Diffusion: Provides rich spatiotemporal dynamics and spatial computing (e.g., maze solving) but requires continuous energy input (dissipative) and faces scalability issues in programmability.
• Synthetic Biology (Organoids/Cells): Exhibits true plasticity and self-healing but suffers from biological variability, slow growth times, and complex maintenance.
• Memristive/Ferroelectric (In-Memory): The most "engineering-ready" solid-state option, offering dense vector-matrix multiplication and reduced data movement. However, it requires algorithm-hardware co-design to handle drift, noise, and non-idealities.
• Photonic: Excels in ultra-high bandwidth and low-latency linear operations (speed of light) but struggles with compact nonlinearity and opto-electronic interface costs.
• Mechanical/Metamaterials: Enables "structural intelligence" and embodied control without electricity, ideal for soft robotics, but limited by the speed of sound and material fatigue.
• Microfluidic/Iontronic: Bridges the gap between electronics and biology, offering direct bio-integration and fluidic logic, though constrained by low bandwidth (kHz) and fabrication complexity.

The paper presents a "Computational Impedance Matching" concept: the value of a substrate emerges when its physical dynamics naturally match the problem's physics (e.g., chemical inputs for chemical networks, RF signals for photonics).

5. Significance and Claims

The paper claims that the proliferation of physical neural substrates represents a shift toward treating physical dynamics as a first-class computational resource, rather than merely an implementation detail.

• Beyond Silicon: The authors argue that exploring alternative substrates is not just about energy efficiency but about enabling capabilities impossible with conventional electronics, such as in-situ molecular decision-making, thermodynamic optimization, and direct bio-integration.
• Co-Design Necessity: A central claim is that successful physical intelligence requires algorithm-hardware co-design. The computational model must be chosen with the substrate's physics in mind (e.g., embracing locality, tolerating imprecision, exploiting native dynamics).
• Modest Expectations: The paper explicitly states that PNNs are unlikely to displace digital computing as a universal platform. Instead, their significance lies in specialized niches where their physical properties provide decisive advantages (e.g., edge sensing, soft robotics, bio-hybrid interfaces).
• Standardization Gap: The work highlights the critical need for standardized reporting and benchmarking to enable cumulative progress and meaningful comparison across such disparate technologies.

In conclusion, the paper positions Physical Neural Computing not as a replacement for digital AI, but as a complementary paradigm that expands the design space of intelligence by embedding computation directly into the matter of the physical world.