Mechanical Neural Networks

Updated 28 January 2026

Mechanical Neural Networks are physical systems where mechanical elements such as springs, masses, and beams directly encode neural weights and dynamics.
They utilize architectures ranging from static spring lattices to dynamic adaptive networks that implement in situ backpropagation for training efficiency.
Practical implementations, including 3D-printed structures and educational demonstrators, reveal advantages like embodied computation, robustness, and self-healing capabilities.

A Mechanical Neural Network (MNN) is a physical, material-based instantiation of neural network computation in which mechanical elements—typically masses, springs, elastic beams, or non-linear mechanical units—directly encode the weights, architecture, and even internal dynamics of a machine learning model. Rather than simulating a neural network in software, an MNN directly leverages the physics of force balance, elasticity, and dynamic response to process inputs and perform learning, in some cases even carrying out adaptation and optimization in situ within the mechanical substrate itself. MNNs span realizations from spatially-distributed spring-mass networks and topological lattices, to adaptive materials, bistable chambers, and analogues of multilayer perceptrons constructed from tangible mechanical linkages. This class of systems forms a subset of physical neural networks (PNNs), situated at the intersection of intelligent metamaterials, embodied computation, and neuromorphic hardware.

1. Fundamental Principles and Network Architectures

The canonical MNN architecture is a network (graph) of nodes connected by mechanical elements—typically linear or weakly-nonlinear springs, beams, or more elaborate units such as bistable elements, adaptive directed springs, or custom linkages. The physical analogues of neural weights are the tunable stiffnesses $k_{ij}$ (or, in some models, the masses $m_i$ , damping coefficients, or resistance values) associated with each connection. The architecture can be lattice-based (1D chains, 2D/3D lattices, honeycomb/metamaterial frameworks), fully connected, or more abstract/disordered depending on the task (Li et al., 2024, Chen et al., 2024, Li et al., 10 Mar 2025, Patil et al., 2023, Ben-Haim et al., 2024).

In static MNNs, input signals are applied as external forces or displacements at prescribed “input” nodes, while outputs are encoded as the steady-state displacements (activations) $u_i$ at designated “output” nodes. The equilibrium of the structure is governed by force balance equations: $D(k) u = F,$ where $D(k) = C^T K C$ is the global stiffness matrix (with $K$ diagonal in the $k_i$ , and $C$ the incidence/compatibility matrix), $F$ is the vector of nodal external forces, and $u$ the vector of nodal displacements.

Extensions to dynamic MNNs introduce inertia: the response to time-varying or oscillatory inputs includes mass-augmented dynamics, with equations of motion involving the dynamical matrix, and outputs emerging as modal energies, frequencies, or time-domain deflections (Li et al., 10 Mar 2025, Chen et al., 2024).

Nonlinear and adaptive MNNs may employ elements exhibiting geometric or material nonlinearity, bistability, or even active adaptation, such as adaptive directed springs (Patil et al., 2023) or multistable networks (Ben-Haim et al., 2024). In these, local responses and learning rules are mediated by the evolving mechanical state.

2. Mechanical Backpropagation and Learning Mechanisms

Training an MNN involves adapting the physical parameters (spring constants, masses, geometrical features) to minimize a loss functional $\mathcal{L}$ that quantifies discrepancy between network output and task target. The mechanical analogue of backpropagation is typically realized via implicit differentiation (adjoint method) of the equilibrium or eigenproblem constraints (Li et al., 2024, Chen et al., 2024, Li et al., 10 Mar 2025):

$\min_k \mathcal{L}[u(k)] \ \text{subject to}\ D(k) u = F$

with

$\frac{\partial \mathcal{L}}{\partial k_i} = u_\text{adj}^T C^T E_i C u = e_{\text{adj},i} \cdot e_i,$

where $u_\text{adj}$ solves the adjoint problem $D u_\text{adj} = -(\partial \mathcal{L}/\partial u)^T$ , $E_i$ selects the $i$ th spring, and $e$ and $e_\text{adj}$ are forward/adjoint bond elongations (Li et al., 2024).

This local gradient enables either:

In silico training: Gradients are computed and used to update the $k_i$ parameters within a simulation or electronic control system (Chen et al., 2024).
In situ training: Gradients are measured physically by applying prescribed forces, measuring displacements/elongations, and locally modulating mechanical parameters (e.g., by tunable-stiffness materials or mechanical actuators), enabling the network to learn directly in the mechanical domain (Li et al., 2024, Li et al., 10 Mar 2025).

For dynamic MNNs, derivatives propagate through eigenvalue sensitivities of the dynamical matrix, allowing gradient descent on modal frequencies, dispersion properties, or classification energies (Chen et al., 2024, Li et al., 10 Mar 2025).

Table 1: Gradient Computation in MNNs

MNN Type	Governing Equation	Gradient Expression
Static spring lattice	$D(k)u = F$	$e_{\text{adj},i} \cdot e_i$
Dynamic mass-spring net	$(-\omega^2 M + D)U = F$	$2\mathrm{Re}[E_{\text{adj},i} E_i]$

3. Experimental Implementations and Physical Realizations

State-of-the-art MNNs have been experimentally demonstrated in a variety of forms:

3D-printed MNNs: Networks of slender beams or bars (cross-sections $\sim2$ mm) in soft polymer matrices (e.g., Agilus30), with nodes fixed at boundaries and input forces applied as hung weights. Bond elongations and nodal displacements are tracked by high-resolution cameras and correlation-based tracking algorithms. In current laboratory implementations, computed gradients are used to redesign and reprint bar widths for next-generation networks; in future, real-time adaptation could be achieved with smart materials (Li et al., 2024).
Physical perceptrons: Hands-on educational MNN demonstrators use levers, clamps, threads, and pulleys to realize multilayer perceptrons (MLPs) with ReLU activations; students physically move clamps to adjust weights and explore logical operators (AND, OR, NOT, XOR) and real-valued function approximation (Schaffland, 2022).
Adaptive mechanical circuits: MNNs constructed from adaptive directed springs utilize energy-harvesting through ratchets and pendulum-gated mechanisms to enable continuous, unsupervised learning from oscillatory environmental stimuli. These implementations are hardware-complete, requiring no electronics or external power input (Patil et al., 2023).
Topological MNNs: Honeycomb spring-mass lattices exploit quantum spin Hall effect (QSHE) analogues for robust, pseudospin-polarized interface modes, enabling classification with resilience to damage. Networks are fabricated via 3D printing or laser cutting, with variable-stiffness springs or tunable geometry for learning (Li et al., 10 Mar 2025).
Multistable/bistable MNNs: Liquid-filled elastomeric chambers networked via viscous resistances exhibit multistability, with distinct equilibrium states encoding digital memory or logic gates. Tubing resistances and morphologies are trained to realize classification, memory, or actuation functionalities (Ben-Haim et al., 2024).

4. Representative Tasks, Demonstrations, and Performance

MNNs have demonstrated wide-ranging task-related competencies, including:

Classification (e.g., Iris dataset): Force/displacement encoding of class features applied to input nodes, read out at output nodes as winner-take-all displacement. In situ backpropagation yields $\sim$ 100% train/test accuracy within 100 epochs in direct physical experiments (Li et al., 2024). Topological MNNs achieve train/test accuracies of 85–96% on multiple datasets, with robustness to bond pruning (Li et al., 10 Mar 2025).
Regression/behavior learning: MNNs are trained to mirror prescribed response functions (e.g., $u_R^y = 0.016 F$ ) over ranges of force input, reaching submillimeter accuracy in both simulation and experiment within a few thousand epochs (Li et al., 2024, Chen et al., 2024).
Morphing/adaptive control: Two-dimensional lattices are optimized to morph under load into controlled shapes (sinusoidal edge contours, target pointer displacements) (Oktay et al., 2023, Chen et al., 2024).
Wave bandgap engineering: MNNs with disordered architecture have been trained to create tunable bandgaps at desired frequencies, forming the basis for deployable sensors and switches (Chen et al., 2024).
Multitask parallelization: Topological MNNs support frequency-division multiplexing, enabling spatiotemporally resolved parallel classification via distinct carrier frequencies assigned to different tasks—realized without architectural duplication (Li et al., 10 Mar 2025).

5. Robustness, Retrainability, and Functional Properties

A defining property of advanced MNNs is their resilience and adaptability:

Self-healing: Experimental removal of critical (vs. redundant) springs can substantially impair performance, but retraining via in situ backpropagation restores significant accuracy. Redundant bonds can be removed without major loss (Li et al., 2024).
Task switching: MNNs can be repurposed between unrelated tasks (e.g., from Iris classification to regression and back), converging in a few hundred epochs to zero loss in both directions. However, the mechanical equilibrium configuration (bar pattern) lands in different local minima for each task (Li et al., 2024).
Topological protection: In TMNNs, the QSHE interface mode topology ensures that even after learning-induced disorder or bond removal, the protected interface state persists, yielding classification robustness far beyond conventional spring-mass lattices (Li et al., 10 Mar 2025).
Continuous online learning and memory: Some MNNs (notably those with ratchet-driven adaptive springs or bistable chambers) can continually adapt to stimuli, store non-volatile memory states, and execute history-dependent transitions, all within the mechanical substrate (Patil et al., 2023, Ben-Haim et al., 2024).

6. Broader Context, Limitations, and Applications

MNNs represent a paradigm shift in neuromorphic computation, moving “learning” into the very fabric of materials and metamaterials. Their principal attributes include:

Local, energy-efficient, and physically interpretable computation: All learning rules are local in the sense that each spring or adaptive element needs only its own measured quantities (e.g., forward/adjoint elongations) for parameter updates (Li et al., 2024, Li et al., 10 Mar 2025, Patil et al., 2023). Weights are “real”—the physical stiffness or geometry of the structure—rather than abstract numbers.
Embodied intelligence and material computation: The mechanical substrate both computes and actuates, blurring the boundary between structure and function. Applications are anticipated in soft robotics, adaptive sensors, deployable metamaterials, battery-free smart devices, and autonomous “self-healing” structures capable of learning from the environment (Li et al., 2024, Chen et al., 2024, Patil et al., 2023).
Scalability and challenges: As in digital ANNs, parameter count and mechanical DOFs scale with network size. However, mechanical nonidealities—such as friction, material fatigue, bandwidth limits, and precision of tunable-stiffness elements—can limit performance at large scale. Actively tunable elements (e.g., those enabling negative stiffness or fast adaptation) remain challenging to realize in bulk hardware. Rapid retraining requires distributed actuation and sensing at high temporal resolution (Chen et al., 2024).
Interpretability and control: Direct correspondence between learned weights and measurable mechanical quantities enables deep inspection, parameter identification, and even equation discovery if combined with monitoring and theoretical models.

A plausible implication is that future developments will increasingly exploit fully in situ, autonomous adaptation via novel materials (magnetoactive, phase-change, phototunable, etc.), and bridge the gap to nonlinear and hybrid tasks, further blurring distinctions between computation, memory, and actuation in engineered systems.

7. Educational and Software-Integrated MNNs

Mechanical realizations of neural networks have also served as unique educational platforms. The wooden-lever MNN demonstrator enables students to build, calibrate, and probe multilayer perceptrons with ReLU activations entirely through tangible manipulation. Despite limitations on numerical accuracy and scalability, this approach compellingly demonstrates the role of weights, activations, biasing, and nonlinearity in a physically embodied manner (Schaffland, 2022).

Concurrently, hybrid frameworks—such as neuromechanical autoencoders—jointly optimize digital network controllers (as boundary actuators) and nonlinear elastic morphologies for co-designed mechanical intelligence, leveraging differentiable simulation for end-to-end training (Oktay et al., 2023). These approaches facilitate transition from simulation to real-world prototypes and provide pipelines for co-design of intelligent structures.

Key References:

"Training all-mechanical neural networks for task learning through in situ backpropagation" (Li et al., 2024)
"Intelligent mechanical metamaterials towards learning static and dynamic behaviors" (Chen et al., 2024)
"Topological mechanical neural networks as classifiers through in situ backpropagation learning" (Li et al., 10 Mar 2025)
"Self-learning mechanical circuits" (Patil et al., 2023)
"The Mechanical Neural Network(MNN) -- A physical implementation of a multilayer perceptron for education and hands-on experimentation" (Schaffland, 2022)
"Neuromechanical Autoencoders: Learning to Couple Elastic and Neural Network Nonlinearity" (Oktay et al., 2023)
"Multistable Physical Neural Networks" (Ben-Haim et al., 2024)