Which Tunnel-Barrier Defects Can Transport Curves Reveal?

A magnetic tunnel junction is a sandwich of two magnets separated by an insulating layer about a nanometre thick. Electrons cross that layer by quantum tunnelling, and the device's resistance depends on whether the two magnets point the same way or opposite ways. Real insulating layers contain defects: impurity atoms, vacancies, and disordered interfaces. These defects create "stepping-stone" energy levels that electrons can hop through, and those hops change how conductance and magnetoresistance vary with temperature and voltage. Physicists usually diagnose barriers by fitting measured curves to one-dimensional formulas that ignore defects, or to phenomenological hopping terms that lump all defects into one rate. Nobody can yet say, from a set of transport curves, where the traps sit, how deep they are, or whether an anomaly comes from the barrier bulk or from the magnetic interface.

We propose to treat defect diagnosis as an inverse problem of the kind the image-reconstruction community already solves with multi-part loss functions. A neural network maps measured conductance-versus-temperature, conductance-versus-bias, and parallel/antiparallel curves to a posterior distribution over a compact defect description: trap density by depth and energy, a spin-flip fraction, an interface-disorder parameter, and effective barrier height and width. Its training loss has two parts. The first compares predictions to simulated ground truth. The second passes the prediction back through a differentiable transport simulator and checks agreement with every measured channel, with each channel weighted by its own measurement noise rather than by a hand-tuned coefficient. The more important deliverable is an identifiability map: which defect parameters are actually determined by which measurements, and which extra measurement breaks the bulk-versus-interface ambiguity. Without that map, any reconstruction risks confidently inventing traps.

Core Claim

Macroscopic transport curves pin down only a few combinations of defect parameters. A reconstruction trained only against ground-truth parameters on simulations will regress toward average barriers and will not stay faithful to the measurements when real devices differ from the simulator. Adding a forward-consistency term, with channels weighted by their noise, fixes the second problem. An explicit identifiability analysis of the simulator's Jacobian tells us which outputs to report and which to withhold. The concrete, falsifiable prediction concerns disentangling two anomalies. The non-monotonic 80–160 K rise in magnetoresistance seen in unannealed junctions should be attributed to bulk traps. The 70 K peak in junctions with a disordered free layer should be attributed to the interface. This attribution should hold only when bias-dependent and antiparallel-state data are included, and it should become ambiguous when they are withheld.

Grounding in Prior Work

The physics side motivates the gap. Temperature dependence of the tunneling magnetoresistance for tunnel junctions reports CoFe/Al₂O₃/CoFe junctions whose magnetoresistance rises between 80 and 160 K before annealing. Annealing at 200 °C removes this behaviour and raises the ratio from 20.7% to 59.0% at 100 K. The authors explain it with a spin-dependent two-step tunnelling model they themselves call phenomenological. Temperature dependence of magnetoresistance in magnetic tunnel junctions with different free layer structures compares junctions that share a pinned layer but differ in free layer. The authors fit Brinkman–Dynes–Rowell (BDR) trapezoidal barriers, extracting 11.3–12.7 Å and 1.28–1.03 eV for Type 1 and 6.7–7.2 Å and 3.46–3.28 eV for Type 2. Type 1's 70 K peak is read as spin-glass-like interface disorder. These two papers show two superficially similar anomalies with different claimed origins, and no shared method for telling them apart. That is the problem we target.

The machine-learning side supplies the method. A Novel Loss Function Incorporating Imaging Acquisition Physics for PET Attenuation Map Generation using Deep Learning pairs an image-domain loss with a loss computed after a differentiable physical projector, cutting projection-domain error by 16.3%. We reuse that structure, with a transport simulator in place of the Radon projector. SINCO adds a downstream-task term (segmentation Dice) to a reconstruction loss. We borrow the idea of a diagnostic head, but constrain how it is used, for the reason below. Lossy Image Compression with Conditional Diffusion Models exposes a continuous knob between fidelity and learned prior. We instead sample from a posterior, so that measurement-consistent diversity is visible rather than averaged away.

One result constrains the design. On The Classification-Distortion-Perception Tradeoff proves that when the degradation is known, restoration cannot lower classification error below what a classifier achieves on the raw degraded signal. Translated here, reconstructing a defect map and then labelling it "bulk" or "interface" cannot beat a classifier trained directly on the raw curves. We therefore treat the raw-curve classifier as a ceiling. The diagnostic term in our loss only asks the reconstruction to agree with that classifier; we do not claim it adds information.

On the simulation side, Deep Learning-Based Quantum Transport Simulations in Two-Dimensional Materials and Deep learning accelerated quantum transport simulations in nanoelectronics predict tight-binding Hamiltonians from atomic positions and feed them to a Green's-function transport solver. This runs 100–1000× faster than first-principles transport, for example 56 s versus 3024 s for graphene. That work runs only forward, has not been applied to magnetic junctions, and trains on single-objective regression. We use it only as an optional higher-fidelity generator for testing model mismatch; it is not our inversion engine.

Low Resistance Asymmetric III-Nitride Tunnel Junctions Designed by Machine Learning fits gradient-boosted trees to device-simulator resistance, reaching 90.5% test accuracy. It is a forward scalar regressor over composition and says nothing about traps.

Research Program

The first component is a differentiable forward model. The barrier is a one-dimensional tight-binding chain between spin-polarised leads. Trap sites are placed in depth-by-energy bins with occupancy given by $\theta$. Coherent transmission comes from the standard Green's-function trace formula. A sequential two-step hopping channel through traps is governed by thermally activated rate equations, and interface disorder enters as a temperature-dependent reduction of interface spin polarisation with a freezing scale. The simulator outputs parallel and antiparallel conductance versus temperature, the bias-dependent differential conductance at several temperatures, and the magnetoresistance ratio. It is written in an autodiff framework so gradients flow to $\theta$.

The second component is the identifiability audit. Across the prior over $\theta$, we compute the Jacobian of the simulator with respect to the parameters and its singular spectrum. We do this for each measurement protocol: temperature sweeps alone, plus bias sweeps, plus antiparallel data, and plus narrower or wider temperature windows. Parameter directions with negligible singular values are declared unidentifiable. The network reports them only as posterior spread and never as point estimates.

The third component is the amortised inverse model. A conditional generator produces posterior samples $\hat\theta$ from measured curves $y$. It is trained on simulated pairs with the loss

$$ \mathcal{L} = \mathbb{E}\,\|\hat\theta - \theta\|^2_{W} \;+\; \sum_{c}\frac{1}{2\sigma_c^2}\,\big\|F_c(\hat\theta) - y_c\big\|^2 \;+\; \gamma\,\|\rho(\hat\theta)\|_1 . $$

Here $W$ restricts the parameter term to identifiable directions from the audit. $F_c$ is the simulator's output for measurement channel $c$, $y_c$ the measured curve, and $\sigma_c$ that channel's noise level, estimated from repeated sweeps. $\rho(\hat\theta)$ is the trap-density part of the prediction, and $\gamma$ is a single sparsity weight. Setting channel weights from noise variances turns the usual brittle weight-tuning into a likelihood choice. The only free weight is $\gamma$, and we report results across a sweep of it. A small diagnostic head that predicts bulk-versus-interface agreement with the raw-curve classifier is added as an optional term and ablated.

Baselines

The naive reference point is the prior mean: predict the average defect configuration for every device. Any method must beat it on parameter error in identifiable directions. It must also beat a single-component network trained only on the parameter term, which is our main ablation and directly tests the forward-consistency ingredient.

The standard physics baseline is a BDR trapezoidal fit to bias-dependent conductance, run as in Temperature dependence of magnetoresistance in magnetic tunnel junctions with different free layer structures. Its reported values, 6.7–7.2 Å and 3.46–3.28 eV for Type 2 junctions, are the reference our effective-barrier outputs must reproduce within uncertainty on clean devices. We pair it with the modified-Julliere two-step fit of Temperature dependence of the tunneling magnetoresistance for tunnel junctions. No reported fit residual or held-out prediction error was found on Lacuna for either model. We will therefore refit both ourselves on identical digitised or re-measured curves, with the same held-out temperatures. Beating these baselines means lower error on withheld temperature and bias points, together with a correct bulk/interface attribution that they cannot express.

The strongest competitor is per-device regularised inversion using our own simulator. This means maximum a posteriori optimisation of the same likelihood-plus-sparsity objective, and Markov chain Monte Carlo sampling of the posterior, with no neural network. It gets the same simulator, the same noise model, and a compute budget matched to the network's training plus inference cost amortised over the test set. Its hyperparameter search gets as many trials as ours. The network "wins" only if it matches this baseline's parameter error and posterior coverage while being much faster, or beats it under simulator mismatch. If it merely matches, we say so.

A direct-regression baseline in the spirit of Low Resistance Asymmetric III-Nitride Tunnel Junctions Designed by Machine Learning uses gradient-boosted trees from summary features of the curves to defect parameters. It is trained on the same simulated set. The raw-curve bulk/interface classifier serves as the ceiling for the diagnostic task, as argued above.

Why not just…?

Why not just fit BDR plus a two-step hopping term, as experimentalists already do? These fits have a handful of lumped parameters. They can reproduce a temperature curve, but they have no variable for trap depth, position, or interface freezing. They therefore cannot say whether Type 1's 70 K peak and the unannealed 80–160 K rise share a cause. The held-out-temperature test and the annealing test below decide whether the extra description buys predictive power or only flexibility. If the lumped fits predict withheld points equally well, the contribution reduces to the identifiability map, and we will report that.

Why not just run regularised least squares or MCMC on each device with the same simulator? This is a real threat: it optimises essentially the same objective without a learned model. On clean simulated data we expect it to match or beat the network in accuracy. The network's case rests on speed across many devices and on robustness when the simulator is wrong. Robustness is tested by inverting data generated with three-dimensional random trap placements or tight-binding Hamiltonians the one-dimensional model lacks. If MCMC wins there too, we adopt it as the inversion engine. The loss design and identifiability audit remain the result.

Why not just train a network with a plain squared error on parameters? That is our single-component ablation. The CDP analysis in On The Classification-Distortion-Perception Tradeoff predicts it will output averaged barriers in unidentifiable directions. The PET result of Shi et al. shows the analogous failure of image-only losses to respect measurements. We test this directly by measuring each method's residual after re-simulating its prediction.

Why not just compensate in hardware, as the dynamic read-bias circuit of A Read Margin Enhancement Circuit With Dynamic Bias Optimization for MRAM does? That circuit cuts bit error rate by one to two orders of magnitude, but it treats the junction as a black box. It solves a different problem, keeping chips working, whereas we want to tell process engineers which defect to remove.

Proposed Experiments

The synthetic benchmark draws about $10^5$ defect configurations from the prior, simulates all channels with realistic noise, and holds out a mismatched test set from the richer generators described above. Metrics are error in identifiable parameter directions, re-simulated measurement residual per channel, credible-interval coverage, and bulk/interface area under the ROC curve relative to the raw-curve ceiling.

The first decisive check is identifiability-driven. The audit predicts which protocol breaks the bulk/interface degeneracy. Attribution accuracy should collapse toward chance when that channel is withheld and recover when it is restored. If attribution works equally well from temperature sweeps alone, our identifiability analysis is wrong.

The second check uses the published CoFe/Al₂O₃/CoFe annealing pair. Inverting the pre-anneal and post-anneal curves separately, the inferred bulk trap density must fall sharply after annealing. The inferred interface parameter should stay roughly unchanged, and the re-simulated magnetoresistance at 100 K should land near the reported 20.7% and 59.0%.

The third check uses the Type 1 and Type 2 free-layer junctions. Type 2's effective barrier should agree with the BDR values within posterior width. Type 1's 70 K peak should be assigned to the interface parameter rather than to bulk traps; the reverse outcome falsifies the attribution.

Across both real datasets, every method is fit on a subset of temperatures and scored on predicting the withheld ones. This is the one metric where all baselines compete on equal footing. Real-data curves will be digitised from the published figures unless raw data can be obtained, and the digitisation noise will set $\sigma_c$.