Reading Strain and Temperature Apart from One RF Magnetoelastic Film

Magnetoelastic thin-film strain sensors are extremely sensitive. A film of amorphous FeCoSiB, a soft magnetic alloy whose magnetic permeability changes sharply under stress, is driven with a radio-frequency current. At these frequencies the current crowds into a thin surface layer, the "skin depth", whose thickness depends on permeability. A tiny strain therefore changes the electrical impedance a great deal. Reported gauge factors reach 1,700–1,900 at 50–100 MHz, about a thousand times a metal foil gauge, and the resolution is about 0.1 microstrain (Highly Sensitive Strain Sensors with FeCoSiB Amorphous Films.). The weakness sits in the same place as the strength. Permeability also depends on temperature, so a warm afternoon and a real bend produce the same shift in impedance when it is read at one frequency.

This note proposes to sweep the carrier frequency and treat the whole complex impedance spectrum as the measurement. A single network with a shared encoder would then be trained to output strain and temperature together. Temperature labels come from a thermistor present only during calibration, so the deployed sensor needs no second device. The bet is that strain and temperature leave different fingerprints across frequency. Temperature also changes electrical conductivity, and permeability relaxes differently at different frequencies. The second bet is that supervising the temperature output forces the shared encoder to separate the two causes. That separation should make strain estimates hold up at temperatures the model never saw in training. Both bets can fail, and the program is built so that each failure is detected cheaply and early.

Core Claim

The multi-frequency impedance spectrum of one FeCoSiB meander film is informative enough to recover strain and temperature jointly without a reference sensor. Joint strain–temperature training beats three simpler options at strain accuracy under temperature change it has not seen: a strain-only model on the same spectrum, a linear inversion of the spectrum, and hardware temperature compensation.

The physical reason to expect separability is written in the skin-depth relation $\delta = \sqrt{2/(\omega\mu\sigma)}$. Here $\omega$ is the carrier angular frequency, $\mu$ the transverse permeability, and $\sigma$ the conductivity. Strain acts almost entirely through $\mu$. Temperature acts through both $\mu$ and $\sigma$, and the frequency dependence of $\mu$ itself shifts as domain-wall motion gives way to moment rotation at higher frequencies (Micromagnetic Simulation of Saturation Magnetization of Nanocrystalline Soft Magnetic Alloys under High-Frequency Excitation). The model can be linearized around an operating point. At carrier frequencies $\omega_1,\dots,\omega_K$, the real and imaginary impedance shifts stack into a vector:

$$\Delta \mathbf{z} = \mathbf{a}\,\Delta\epsilon + \mathbf{b}\,\Delta T + \mathbf{n},$$

In this equation, $\mathbf{a}$ and $\mathbf{b}$ are the per-frequency sensitivities to strain $\epsilon$ and temperature $T$, and $\mathbf{n}$ is readout noise. A single frequency gives one equation in two unknowns, which is the degeneracy described above. The two causes can be separated exactly when $\mathbf{a}$ and $\mathbf{b}$ are far from parallel. How far from parallel they are is measured by the condition number of the matrix $[\mathbf{a}\ \mathbf{b}]$, estimated from calibration data. If that number is huge at every operating point, no learning method can help and the project should stop.

Relation to Prior Work

The hardware papers supply the sensor and motivate the problem. Two papers bonded bias-stress-annealed FeCoSiB films to PZT piezoelectric actuators and fed the single-frequency impedance into an analog feedback loop. That loop linearized the actuator up to 220 Hz and cut head–disk contact vibration from 72 mV to 18 mV (Application of magnetoelastic strain sensors to smart actuators., Linear Drive of Piezoelectric Actuator with Highly Magnetostrictive Film Strain Sensor.). Another paper moved the film to polyimide and reached a figure of merit near 1,500 at 700 MHz (Magneto-elastic Strain Sensor Formed on a Flexible Substrate.). All of these collapse the readout to one scalar at one frequency and leave thermal drift unaddressed. This proposal keeps their devices and changes only what is measured and how it is inverted.

The polyimide work also constrains the design. Its sensitivity comes from residual stress engineered through the thermal-expansion mismatch between film and substrate. Designing temperature sensitivity out of the hardware would therefore also remove much of the strain sensitivity. That is why a computational decoupling is worth attempting.

Two machine-learning lines supply methods. Computational intelligence for soft strain sensor sustainability compensates a low-frequency piezoresistive elastomer sensor with a staged pipeline: a sigmoid for nonlinearity, recurrent models for hysteresis, and a transformer for drift. That pipeline extended usable life from 6 to more than 20,000 cycles. It works on DC resistance, with a single output and staged models. We work on RF spectra, with two physical outputs and one shared encoder, and we use its staged design as a baseline. A Foundation Model for Virtual Sensors shows that joint prediction across related physical channels lowered error on automotive telemetry from 0.177 to 0.121 MSE compared with isolated models. Its attention-based pruning of irrelevant inputs inspires our frequency-selection step. It never touched electromagnetic impedance, and it works at 10 Hz telemetry rates, not RF spectra.

RotoGrad resolves conflicting task gradients by rotating shared features, but it was tested only on vision benchmarks. Our targets have wildly different physical units, microstrain and kelvin. We treat RotoGrad as an optimizer to compare against, not as the contribution, because standardizing the targets may already remove most of the scale imbalance.

Research Program

The first stage is a measurement campaign, since no public dataset of this kind appears on Lacuna. We fabricate FeCoSiB meander films in two forms. One is on glass cantilevers following the geometry in Highly Sensitive Strain Sensors with FeCoSiB Amorphous Films.. The other is on field-annealed polyimide following Magneto-elastic Strain Sensor Formed on a Flexible Substrate.. Each is mounted in a temperature chamber.

A vector network analyzer records complex impedance at roughly 64 log-spaced frequencies from 1 MHz to 1 GHz. Strain is imposed by a PZT actuator or a mechanical bender. Reference strain comes from a laser displacement sensor and a co-located foil gauge. Reference temperature comes from an NTC thermistor calibrated with residual compensation, the method that reached 0.1 K accuracy in Design and Application of High-Precision Temperature Measuring Instrument for Ice Cover Profile of River Based on the Resistance Residual Compensation Method. Conditions follow a grid-plus-random protocol over strain (0 to 900 microstrain) and temperature (a wide band, for example 10–70 °C), with loading and unloading legs and slow thermal ramps. Several devices come from separate deposition runs, so that batch-to-batch variation is measurable.

The second stage estimates identifiability before any network is trained. We fit the local sensitivities $\mathbf{a}$ and $\mathbf{b}$ at each operating point and map the condition number across strain, temperature, and frequency band. This map is a result in its own right. It tells hardware designers which carrier bands carry temperature-specific information.

The third stage is the model. A small encoder, either a one-dimensional convolution over frequency or an MLP, maps the spectrum to a shared representation. It optionally takes a short history of past spectra so that it can capture hysteresis. Two heads read strain and temperature from this representation. Training minimizes

$$\mathcal{L} = \mathcal{L}_\epsilon + \lambda\,\mathcal{L}_T,$$

where both are mean-squared errors on standardized targets and $\lambda$ weights the temperature output. At deployment only the spectrum is needed. A sparsity penalty on a learned per-frequency gate then finds the smallest set of carrier frequencies that keeps accuracy. This matters because a practical readout with three or four fixed oscillators is far cheaper than a network analyzer, and a short sweep also preserves bandwidth for dynamic strain.

Baselines

The naive reference point is the standard practice in the cited hardware papers: a linear calibration of impedance magnitude at one frequency (100 MHz) against strain, fit at room temperature. Its error under temperature change quantifies the problem. On this question, no reported result was found on Lacuna for thermal apparent strain in FeCoSiB RF sensors. The only related published figure is the 46% impedance change at 300 microstrain at 100 MHz from Highly Sensitive Strain Sensors with FeCoSiB Amorphous Films.. We will therefore measure this baseline ourselves.

The hardware-compensation baseline gives the same single-frequency calibration a deployed temperature reference. In one version, a thermistor feeds a fitted temperature-correction term. In the other, an unstrained dummy film on the same substrate is read in a differential configuration. This is what an engineer would do first. It uses extra hardware, which the proposal tries to avoid, so matching it without the extra hardware already counts as success.

The strongest competitor is linear inversion of the full spectrum. This covers least-squares inversion of the equation above using calibrated $\mathbf{a}$ and $\mathbf{b}$, and ridge or partial-least-squares regression from all 64 complex features to strain and temperature. If the physics is close to linear over the working range, this is all one needs. No reported result was found on Lacuna for it on magnetoelastic impedance spectra.

The ablation that isolates the key ingredient is a strain-only network. It has the same architecture, data, and temperature-varied training set, but no temperature head ($\lambda = 0$). It could in principle learn temperature invariance from data alone. For hysteresis, we also run two further baselines on the same data. One is SINDy through the SmartHysteresis code, which reached $R^2 = 0.99$ on 45,000 points of experimental piezoelectric cantilever data. The other is the staged sigmoid → recurrent → transformer pipeline of Computational intelligence for soft strain sensor sustainability. Finally, we compare fixed $\lambda$, uncertainty-style weighting, and RotoGrad, to find out whether gradient conflict between physically dissimilar targets actually arises.

Every method sees identical training splits and the same input features, and none uses the thermistor at test time except the hardware baseline. Each learned method gets the same hyperparameter budget of 50 random-search trials, with selection on a validation split from within the training temperature band. The joint model counts as beating the full-spectrum linear baseline only if it lowers strain RMSE on held-out temperatures by at least 25%, with the gain holding across all devices under device-level bootstrap. It also needs a smaller slope of apparent strain per kelvin. A smaller margin would be reported as "linear inversion suffices".

Why not just…?

Why not just add a thermistor or dummy film and subtract? This works when the temperature at the reference matches the temperature in the film, and it is our hardware baseline. Two problems remain. It adds a component and thermal mass, which is what flexible and embedded deployments want to avoid. It also assumes the temperature effect adds cleanly on top of the strain effect. On polyimide, temperature changes the built-in residual stress that sets the strain sensitivity itself (Magneto-elastic Strain Sensor Formed on a Flexible Substrate.), so the two effects interact rather than simply adding. The comparison against the hardware baseline on polyimide devices will show whether this interaction matters in practice.

Why not just invert the spectrum linearly? This is the real threat. If the identifiability map shows well-conditioned, roughly constant sensitivities over the working range, linear least squares will be close to optimal and the network adds nothing. We will report that outcome as the finding. It would still be new and useful: a two-frequency analog readout that decouples temperature. The network earns its place only where sensitivities vary with the operating point, near the peak of maximum sensitivity whose location shifts with annealing and temperature. The held-out-temperature test is designed to expose exactly that regime.

Why not just train a strain-only network on temperature-varied data? It sees the same spectra and might learn invariance by itself. Our hypothesis is that it learns invariance only inside the range of temperatures it saw in training, whereas a temperature head forces the encoder to represent temperature explicitly. An explicit temperature representation should keep strain estimates stable outside that range. The ablation at $\lambda = 0$, evaluated in and out of the training temperature band, settles this directly.

Why not just use SINDy or a Bouc–Wen-type hysteresis model? SINDy is fast and interpretable, reaching $R^2 = 0.99$ with training times of 0.018 s (Discovery of sparse hysteresis models for piezoelectric materials). Its authors note that it is sensitive to outliers, which RF readouts produce through interference spikes. It also identifies one input–output system at a time and offers no route to a second, unobserved cause. We therefore use it where it fits, on the hysteresis of a known-temperature trace, and compare it on the loading/unloading gap metric. We do not expect it to handle the decoupling problem.

Proposed Experiments and Decisive Checks

The first decisive check is identifiability. If the median condition number of $[\mathbf{a}\ \mathbf{b}]$ implies that estimated temperature carries more than about 1 K of uncertainty at our noise floor across every frequency band, reference-free decoupling is not viable for this film. In that case we report which physical parameter to change, such as film thickness relative to skin depth or the meander geometry.

The second check is temperature extrapolation. We train on the middle of the temperature band and test on both ends, reporting strain RMSE in microstrain, temperature RMSE in kelvin, and apparent strain per kelvin for every baseline. The third check is transfer across devices. We train on films from some deposition runs and test on another run, with and without a few minutes of recalibration data. This tests whether the shared representation captures physics or memorizes one device.

The fourth check is frequency sparsity. We find the smallest set of carrier frequencies that stays within 10% of full-spectrum accuracy. That set is then validated with a fixed-frequency readout on the PZT actuator from Application of magnetoelastic strain sensors to smart actuators., with temperature varying during the test. The decisive outcome there is whether actuator linearity is maintained under a thermal ramp that breaks the single-frequency loop.

Each check has a defined negative result. The linear baseline winning, or the strain-only network matching the joint model, would each be reported as the answer, not argued away.