Mechanical Backpropagation in Dynamic Lattices

• Mechanical backpropagation is a method of implementing gradient-based learning directly within mechanical networks and metamaterials using local physical interactions.
• It utilizes adjoint methods and implicit differentiation to compute exact gradients from nodal displacements and bond elongations.
• Applications include regression, classification, and bandgap control in both simulation and hardware, demonstrating scalability and robustness.

Mechanical Backpropagation is the class of algorithms and physical design principles that realize gradient-based learning, specifically the backpropagation of errors, directly and locally in mechanical systems—most commonly in networks of masses and springs (mechanical neural networks, or MNNs), mechanical metamaterials, and wave-based mechanical lattices. These approaches enable both in silico (simulation-based) and in situ (hardware-based) optimization of mechanical system parameters for machine learning and inverse design tasks. The emergence of mechanical backpropagation provides a rigorous framework for embedding learning and adaptation directly within physical substrates, leveraging adjoint methods, implicit differentiation, and physical reciprocity to enable exact and efficient computation of gradients at the hardware level (Li et al., 2024, Li et al., 10 Mar 2025, Chen et al., 2024).

1. Mechanical Neural Network Architectures

Mechanical neural networks are typically realized as spatially extended networks of point masses (nodes) interconnected by linear springs (edges) whose stiffnesses are the learnable parameters. The mechanical state of the system is represented via node displacements $u \in \mathbb{R}^{dn}$ (for $n$ nodes in $d$ dimensions), with the network's connectivity encoded in a compatibility matrix $C \in \mathbb{R}^{m \times dn}$ (mapping node displacements to spring elongations). The system's static and dynamic behavior is governed by the equilibrium (or dynamical) equation: $D u = F,$ where $D = C^\top K C$ is the stiffness (or dynamical) matrix with $K = \operatorname{diag}(k_i)$ the diagonal matrix of spring constants and $F$ the external nodal force. For forced harmonic (dynamic) response, the governing equation generalizes to

$(-\omega^2 M + D) U = F,$

where $M$ is the (diagonal) mass matrix and $U$ is the vector of complex mode amplitudes (Li et al., 2024, Li et al., 10 Mar 2025, Chen et al., 2024).

Mechanical design variants include:

• 1D and 2D lattices: Serial chains or planar triangulated/honeycomb networks, allowing general static or vibrational behavior (Chen et al., 2024, Li et al., 10 Mar 2025).
• Topological mechanical neural networks (TMNNs): Architectures with QSHE-like band topology, supporting robust pseudospin edge modes for classification (Li et al., 10 Mar 2025).
• Disordered and bandgap-engineered lattices: Lattices with tailored wave response via spatial disorder and targeted eigenvalue control (Chen et al., 2024).

2. Local Gradient Computation via Adjoint and Implicit Differentiation

The fundamental principle underpinning mechanical backpropagation is that the partial derivative of a loss functional $\mathcal{L}$ (with respect to a parameter $k_i$ such as a spring constant) can be computed via an adjoint field, itself derived physically through a complementary experiment or measurement. For the static case, if $\mathcal{L} = \mathcal{L}(u(k))$, then

$$\frac{d\mathcal{L}}{dk_i} = \sum_{a=1}^{dn} \frac{\partial\mathcal{L}}{\partial u_a} \frac{\partial u_a}{\partial k_i}.$$

Given $D u = F$, implicit differentiation gives

$$\frac{\partial u}{\partial k_i} = -D^{-1} \frac{\partial D}{\partial k_i} u,$$

so

$$\frac{d\mathcal{L}}{dk_i} = u_{\mathrm{adj}}^\top \frac{\partial D}{\partial k_i} u,$$

where $u_{\mathrm{adj}}$ solves $$D u_{\mathrm{adj}} = -(\partial\mathcal{L}/\partial u)^\top$$ (Li et al., 2024, Chen et al., 2024).

Physically, this means each component of the parameter gradient can be computed as the product of the $i$th bond's forward elongation ($e_i = C_i u$) and its adjoint elongation ($e_{\mathrm{adj},i} = C_i u_{\mathrm{adj}}$): $$\boxed{ \frac{d\mathcal{L}}{dk_i} = e_{\mathrm{adj},i} e_i }$$ This is notable for its strict locality: only measurements at the $i$th spring are required.

In vibrational (dynamic) systems, analogous adjoint wavefields are computed in the frequency domain, yielding local update rules involving real parts of products of bond elongations in forward and adjoint experiments (Li et al., 10 Mar 2025).

3. In Situ and In Silico Backpropagation Schemes

Mechanical backpropagation can be implemented both numerically (in silico) and directly in physical hardware (in situ):

• In Silico: The network is modeled and simulated using ODE or FEM solvers. Differentiation is performed via implicit or automatic differentiation, involving the solution of linear systems for each parameter (Chen et al., 2024). Efficient algorithms exploit network sparsity to scale to large lattices.
• In Situ: All measurement and computation is distributed physically. Each training epoch consists of two physical experiments:
  • Forward Pass: Apply physical input (forces), measure nodal displacements and bond elongations.
  • Adjoint Pass: Apply adjoint forces derived from the loss gradient, measure nodal and bond responses.
  • Local Update: Each parameter is adjusted by a simple (measured) product rule (Li et al., 2024, Li et al., 10 Mar 2025).

In wave-based TMNNs, both the forward and adjoint responses are measured at the relevant frequencies, and classification output is read out via spatially or spectrally localized observables (Li et al., 10 Mar 2025). The in situ paradigm circumvents global matrix inversions, allowing real-time, local learning that is robust to device-level variability and noise.

4. Task Applications: Regression, Classification, and Bandgap Control

Mechanical backpropagation provides a systematic procedure for configuring mechanical systems to perform:

• Static regression and shape-matching: Networks learn to map input displacements or forces onto prescribed output deformations, with learning evaluated by mean-squared error (Chen et al., 2024, Li et al., 2024).
• Supervised classification: Mechanical classifiers assign input feature vectors (encoded as force vectors or displacement patterns) to output classes, either via spatial displacement magnitude or modal participation ratio. TMNNs leveraging QSHE topology achieve classification on standard datasets (Iris, Penguins, Seeds) with $85$–$96\%$ test accuracy (Li et al., 10 Mar 2025, Li et al., 2024).
• Dynamic bandgap engineering: Networks are trained to achieve target vibrational bandgaps or dispersion relations by minimizing losses defined over eigenfrequencies, enabling precise wave propagation control in disordered metamaterials (Chen et al., 2024).
• Retrainability and damage recovery: Mechanical networks can be retrained for new tasks starting from previously learned configurations and can recover performance after structural damage by continued in situ learning (Li et al., 2024).

The table below summarizes characteristic applications and observed performance for selected architectures:

Application	Mechanical System	Performance/Result
Linear regression	1D/2D MNN, static	MSE $\lesssim 10^{-6}$ after training
Classification (Iris)	TMNN	85–100% train/test accuracy
Bandgap engineering	Disordered MNN	Target gaps realized in dispersion
Retraining after damage	Generic MNN	Accuracy recovery $\sim$80%

5. Physical Principles: Locality, Reciprocity, and Topological Robustness

Mechanical backpropagation exploits several foundational physical principles:

• Locality: All gradient computation and parameter updates require only strictly local physical measurements (bond elongations, nodal displacements), making the approach scalable and hardware implementable (Li et al., 2024, Li et al., 10 Mar 2025).
• Adjunction and Reciprocity: The underlying mathematical structure is adjoint-based, treating the network as a reciprocal physical medium. Gradient information is transmitted via physical states under appropriately applied forces (Hermans et al., 2014, Li et al., 2024).
• Topological Protection: TMNNs with nontrivial band topology inherit robustness to disorder and bond pruning, enabling classification tasks to persist even after structural defects, except at critical interface bonds (Li et al., 10 Mar 2025).
• Parallelization and Multiplexing: Frequency-division multiplexing allows simultaneous training or inference on multiple tasks by assigning distinct frequencies to each, leveraging the continuum of topological edge modes (Li et al., 10 Mar 2025).

6. Algorithmic Implementation and Experimental Realizations

An archetypal mechanical backpropagation workflow proceeds as:

1. Prepare the physical network with tunable stiffness elements and appropriate sensing (e.g., high-speed cameras, laser vibrometers, strain gauges).
2. Apply input pattern (force/displacement), measure response.
3. Evaluate loss and compute loss gradients with respect to outputs.
4. Perform adjoint experiment by applying the adjoint force pattern, measure resulting responses.
5. Form local gradients and update parameters (spring constants, masses).
6. Iterate for multiple training epochs, possibly switching tasks or repairing after damage (Li et al., 2024, Li et al., 10 Mar 2025).

Literature reports model-free experimental gradient accuracy exceeding 90% in regression and behavior-shaping tasks, near-100% classification on canonical datasets, and resilient retrainability post-damage (Li et al., 2024). Network size and computational cost scale favorably due to batch updates and local gradient calculations.

7. Broader Theoretical Context and Physical Realizability

Mechanical backpropagation is part of a broader movement to implement learning dynamics as the spontaneous relaxation of physical systems. Recent theoretical advances establish rigorous correspondences between backpropagation and continuous-time physical relaxation, saddle-point dynamics on doubled state spaces, and local wave scattering processes (Scurria, 2 Feb 2026, Pehle et al., 11 Feb 2026). These formalisms show that:

• Exact gradients can be computed in finite time via relaxation processes entirely governed by local physical interactions.
• Wave-propagating substrates (mechanical, optical, acoustic) can support gradient-based learning as a consequence of local dissipation and impedance matching, obviating the need for digital (global) backpropagation (Pehle et al., 11 Feb 2026).

These foundations position mechanical backpropagation as a scalable paradigm for neuromorphic mechanical computing, intelligent material design, and autonomous learning in physical networks (Li et al., 2024, Li et al., 10 Mar 2025, Scurria, 2 Feb 2026, Pehle et al., 11 Feb 2026).