Layerwise Conductance

Updated 8 February 2026

Layerwise conductance is a metric that quantifies the local transport and functional contribution of individual layers in multilayer systems, spanning quantum materials to neural networks.
It employs rigorous measurement and attribution methods, such as layer-resolved spectroscopy and gradient-based techniques, to isolate individual layer contributions.
Insights from layerwise conductance enable optimization of device performance in nanoscale electronics and improve model interpretability in advanced deep learning architectures.

Layer wise conductance refers to the quantification, often with high spatial or architectural resolution, of electronic, thermal, or algorithmic transport or attribution within a multilayer or hierarchical physical system (such as van der Waals heterostructures or deep neural networks). The term has rigorous definitions and measurement methodologies in both condensed matter physics—where it characterizes spatially resolved transport in multilayer structures—and machine learning, where it encapsulates the functional contribution of each layer or block in a deep model towards specific computational objectives. This overview synthesizes results from electronic and phononic layerwise conductance in van der Waals and layered quantum materials, as well as algorithmic "layer conductance" attribution methods in neural networks.

1. Physical Definitions and Measurement Approaches

In electronic and thermal transport theory, "layer wise conductance" refers to the (often local) conductance $\sigma_{i}$ or $G_{i}$ associated with the $i$ -th layer (or layer pair) in a multilayer system. This can be quantified in various bases depending on the experimental arrangement and the nature of disorder, coupling, and stacking:

Electronic Interlayer Conductance: For multilayer graphene, per-layer (or interlayer) conductance $G(\theta, T) = \frac{1}{A}\left.\frac{dI}{dV}\right|_{V \to 0}$ is defined for pairs of layers, with explicit dependence on twist angle $\theta$ and temperature $T$ , reflecting the sensitivity to misalignment and phonon-assisted tunneling (Sun et al., 2016).
Thermal Layerwise Conductance: In few-layer 2D materials, in-plane thermal conductivity $k$ , and interfacial conductance $G_{\rm int}$ are layer-number resolved using methods such as optothermal Raman, with modeling to extract the contribution of each layer and its coupling to substrates (Easy et al., 2020).
Disordered Layered Systems: For weakly coupled layers with planar and bulk disorder, the layerwise out-of-plane (interlayer) conductivity $\sigma_{zz}$ quantifies vertical charge transport and its suppression or enhancement due to localization/delocalization mechanisms (0902.4847).
Quantum Optical and Transport Probes: Layer-resolved dynamical conductivities $\sigma_j^N(\omega)$ can be defined and computed for each layer $j$ in an $N$ -layer stack, forming the basis for reflectance and transport calculations at finite frequency and field (Sasaki, 2020).

2. Theoretical Formulation and Scaling Laws

2.1. Multilayer Graphene and Dimensional Crossover

The conductivity of a multilayer system can be formally decomposed by: $\sigma_N(\omega, T) = N\,\sigma_0 + R(\omega, T)$ where $\sigma_0$ is the universal single-layer conductivity, and $R$ characterizes interlayer corrections due to finite hopping $t_\perp$ and Coulomb interactions. The dimensional crossover energy $E_c$ , which sets the scale where the system transitions from 2D to 3D coherent transport, is renormalized by intra-layer interactions: $E_c \simeq t_\perp \left(\frac{t_\perp}{t}\right)^{\eta/(1-\eta)}$ with $\eta$ the anomalous dimension from the RG flow (Mastropietro, 2011). For $T,\omega \gg E_c$ , each layer acts as an independent conductor.

2.2. Stacking Structure and Layer-Resolved Conductivity

Stacking sequence (AB/Bernal, ABC/rhombohedral, AA/hexagonal) dictates the Landau spectrum, energy scaling, and selection rules for interlayer coupling:

AB stacking: Layer-resolved Hall and longitudinal conductivities, $\sigma_{ij}^{kl}$ , display non-quantized Hall steps and strong suppression of interlayer longitudinal conductivity near charge neutrality. Analytical results show only a finite "tunneling range" in $|k-l|$ for nonzero interlayer response.
AA stacking: N nearly decoupled monolayers give near-independent layer conductances.
ABC stacking: Chiral Hamiltonian structure allows for large zero-mode degeneracies and non-monotonic, layer-dependent response (Wakutsu et al., 2011).

A general layer-resolved Kubo formula includes the overlap of wavefunctions on layers $k$ and $l$ , capturing the spatial decay and possible sign changes of interlayer conductivity.

3. Disorder, Stacking Faults, and Layerwise Conductance Distribution

Planar disorder (e.g., stacking faults) and random twist angles $\theta$ in van der Waals heterostructures induce strong spatial variability in layerwise conductance:

Phonon-assisted tunneling: Interlayer conductance follows

$G(\theta, T) = G_0 \exp(-\alpha\,\theta) n(\omega_0, T)$

with exponential angular suppression and thermally activated phonon occupation (Sun et al., 2016).

Distribution and Yield: Random stacking leads to an exponential-like distribution for total resistance,

$P(R)\propto \exp(-R/R_0)$

and order-of-magnitude variations in $G_i$ for each bilayer. For device design, control of $\theta$ across the stack is essential for reproducibility.

Disorder-Induced Delocalization: In the presence of both planar and bulk disorder, the dc $\sigma_{zz}$ vanishes without bulk scattering (Anderson localization), while finite bulk disorder delocalizes electrons and produces a non-monotonic dependence of $\sigma_{zz}$ on disorder strength and energy (0902.4847).

4. Layerwise Conductance in Artificial Neural Networks

4.1. Formalism for Layerwise Attribution

"Layerwise conductance" in deep models quantifies the internal contribution of each layer or coarse block to a chosen output, generalizing gradient-based attribution methods to hidden activations. The approach is structurally analogous to physical transport models:

Conductance Score: For a model $m$ and input $x$ , block $i$ receives scalar conductance

$g_i^m(x) = \text{Mean}_j\left(|\text{Cond}_{i,j}^m(x)|\right)$

where each $\text{Cond}_{i,j}^m(x)$ integrates the gradient path from a reference (usually zero) between baseline and input, leveraging the Integrated Gradients formalism (Yang et al., 1 Feb 2026).

Task Embedding: For a task $T$ , conductance vectors are averaged over task data and normalized,

$u_{m,T} = \frac{\tilde{U}_{m,T}}{\max(\|\tilde{U}_{m,T}\|_2, \varepsilon)}$

yielding a functionally salient embedding for the task in the conductance space of model $m$ .

4.2. Transferability and Model Selection

Layerwise conductance serves as the basis for advanced similarity and transferability metrics:

Directional Conductance Divergence (DCD):

$D_m(T \rightarrow S) = \sum_{i=1}^{d_m} \alpha_{m,T}(i) \cdot \delta_m(i; T \rightarrow S)$

evaluates how well the source task $S$ covers the most relevant blocks for target $T$ , with $\alpha_{m,T}$ distributing probability mass via softmax-aligned entropy regularization.

Empirical results demonstrate that conductance-based, model-specific, and asymmetric metrics predict zero-shot model rankings for downstream vision-language tasks with substantial gains over text-based or symmetric baselines (Yang et al., 1 Feb 2026).

4.3. Positional and Word-Type Bias in LLMs

Conductance-based frameworks have been used to dissect positional and lexical biases across transformer layers in LLMs. The layerwise positional profile (fractional conductance assigned to each position) is sharply recency-biased in deeper layers, with early layers showing more sensitivity to content words and primacy effects. This mapping is invariant to lexical scrambling, indicating that positional preference is an architectural property rather than a learned semantic association (Rahimi et al., 7 Jan 2026).

5. Experimental and Theoretical Examples Across Materials and Modalities

5.1. Thermal Conductance

Layer-resolved measurements of thermal conductance in few-layer TMDCs, such as WSe $_2$ , reveal a monotonic decrease in in-plane $k$ and a plateau in interfacial $G_{\rm int}$ with layer number—tracing the cross-plane vibrational mode coupling and phonon-phonon scattering (Easy et al., 2020). Similar methodologies underlie the extraction of interface thermal resistance in more generic layered systems (Zhang et al., 2019).

5.2. Quantum Interference and Topological Effects

Quantized, layer-number dependent conductance oscillations emerge in antiferromagnetic layered topological insulator tunnel junctions, where interference patterns shift with even/odd layer parity, resulting in $G(B, N) = G_0[1 + \cos (2\pi \Phi/\Phi_0 + \delta_N)]$ with $\delta_N = 0$ (odd) or $\pi$ (even) (Choi et al., 2023). This effect is rooted in symmetry-protected phase shifts and underpins topological device function.

5.3. Machine Learning Prediction of Layer-Resolved Quantum Transport

Machine learning frameworks (e.g., clusterized Gradient Boosting Regressors) have been trained to map device geometry, energy, and twist angle to full layer-resolved conductance matrices in twisted bilayer graphene, demonstrating predictive accuracy for counterflow, drag, and Hall responses as experimentally measured (Kuhn et al., 17 Feb 2025).

6. Experimental Implications and Device Applications

Switching Devices: Layerwise suppression or enhancement of conductance—controlled via stacking, twist angle, disorder, or external field—enables switching between on/off states in multilayer graphene and van der Waals heterostructures.
Thermal Management: Quantitative knowledge of layerwise and interface conductances enables predictive multilayer device thermal modeling, critical for nanoscale electronics and thermoelectric applications.
Explainability in AI: For deep models, layerwise conductance provides actionable and architecture-aware insights into model selection, transferability, and functional routing, robust to data and semantic perturbations.

7. Limitations, Generalizations, and Open Directions

Layerwise conductance frameworks rely on idealizations—e.g., perfect stacking, absence of additional inelastic or nonlocal effects, exact boundary conditions in Kubo/Keldysh calculations, or the assumption that scalar objectives sufficiently reflect information flow in neural networks. In physical systems, nonlocal corrections (interlayer induced currents from nonlocal fields) can be incorporated via full conductivity matrices, improving agreement with reflectance and transport experiments in real materials (Sasaki, 2020). For transfer learning, the extension of conductance-based embeddings to architectures beyond vision-LLMs or to more complex tasks remains an active area for methodological development (Yang et al., 1 Feb 2026).

In summary, "layer wise conductance" provides a foundational paradigm for quantifying the local, spatial, or hierarchical flow of electrical, thermal, or algorithmic "current" in structured multi-component systems. Its rigorous definitions, analytic and computational frameworks, and empirical validation extend from quantum layered materials physics to modern explainable AI. The layerwise decomposition of conductance not only reveals fundamental transport and functional mechanisms but also underpins practical device engineering and model interpretability across multiple disciplines.