Chemical sensors that work well in the lab often fail in the field. The substance being measured sits inside a messy background, such as blood, edible oil, wastewater, or ambient air, and that background suppresses, shifts, or mimics the signal. Chemists call this matrix interference, and the "matrix" is everything in the sample except the target. Federated learning, where many devices jointly train one model without sharing raw data, looks like a natural way to learn a sensor decoder that holds up across many backgrounds. It runs into a second problem. The devices in a sensor fleet range from edge GPUs down to microcontrollers that cannot run backpropagation at all, and the cheapest devices tend to sit in the harshest settings: field sites, undiluted samples, electrodes that have fouled over time. Standard federated methods cope with weak devices by dropping them or handing them a thinner model. Both responses quietly remove the hardest backgrounds from training.
We propose splitting the sensor model along the line the chemistry already suggests. Each device keeps a small, physically structured "matrix layer": a per-channel gain for signal suppression plus a low-rank baseline for background absorbance or capacitive drift. The device fits this layer in closed form from a few reference or spiked measurements, which amounts to standard addition carried out as a learned layer. The shared decoder sits on top of the corrected signal. Its final regression head is aggregated through additive summary statistics that any device able to run a forward pass can compute. The weakest devices therefore still shape the shared model, and they never backpropagate. The claim to test is simple: making the matrix correction closed-form and keeping it local is what lets a federated sensor fleet include its cheapest and most matrix-afflicted members. Including those devices should cut error on hard backgrounds more than any capacity-scheduling scheme that leaves them out.
Core Claim
Matrix interference has a known, low-dimensional structure in many sensing modes. Ion suppression and electrode fouling mostly scale the sensitivity, which is the slope of signal against concentration. Pigment absorbance, scattering, and double-layer charging mostly add a smooth baseline. The research file documents these mechanisms and their classical remedies: dilution, standard addition, which spikes known amounts of analyte into the real sample so the matrix acts on native and added analyte alike, and selective electrode materials. We write the raw multichannel response of device $k$ in session $t$ as
$$x_{k,t} = G_{k,t}\, s(c) + B\,\beta_{k,t} + \epsilon,$$
where $s(c)$ is the matrix-free response to analyte concentration $c$, $G_{k,t}$ is a diagonal gain matrix capturing suppression or fouling, $B$ is a small shared basis of background shapes with local coefficients $\beta_{k,t}$, and $\epsilon$ is noise. Given a handful of reference measurements, which could be a spike, a calibrant gas, or a blank, $G_{k,t}$ and $\beta_{k,t}$ follow from a small least-squares solve. The device applies the inverse correction $\tilde{x} = G_{k,t}^{-1}(x - B\beta_{k,t})$ and never transmits $G$ or $\beta$.
On top of the corrected signal, a shared feature extractor $\phi$ feeds a ridge-regression head. The server solves for the head weights $w$ exactly from summed statistics,
$$w = \Big(\sum_k \Phi_k^\top \Phi_k + \lambda I\Big)^{-1} \sum_k \Phi_k^\top y_k,$$
where $\Phi_k$ stacks the features $\phi(\tilde{x})$ for device $k$'s labeled samples, $y_k$ holds their reference concentrations, and $\lambda$ is a ridge penalty. These sums are additive and order-free, so a late or intermittently connected device's statistics can simply be added when they arrive, with no staleness correction. Devices with enough memory also update $\phi$ and the basis $B$ by ordinary federated gradient steps. Microcontrollers only run $\phi$ forward and accumulate $\Phi^\top\Phi$ and $\Phi^\top y$. What is new is not the idea of partially local parameters. The new part is choosing the local part so that it is physically interpretable and closed-form, then using that choice to turn device heterogeneity from a reason to exclude devices into a difference in what each device contributes.
Relation to Prior Work
Two papers supply the general pattern and motivate the gap. Federated Reconstruction splits parameters into global and local sets, rebuilds the local set with a few gradient steps each round, and reports MovieLens RMSE of 0.907 against 0.923 for centralized training. Its local reconstruction needs backpropagation on every device, and its local parameters stand for user preferences, not a physical transform of the input. Decoupled Low-Rank Adaptation keeps one adapter matrix local and aggregates the other to separate shared structure from label noise, reaching 88.30% on CIFAR-100 with 40% pairflip noise against 60.11% for plain federated LoRA. It treats noise as discrete label corruption. We borrow its asymmetric shared/local split but apply it to continuous distortion of the input.
The closest work on correcting "site effects" is Distributed Harmonization, which estimates per-site feature means and scales, clusters them on the server, and applies cluster-level affine corrections to multi-site MRI. It is the strongest precedent for affine per-site correction in a privacy setting. It assumes capable, synchronous institutional clients and corrects features after the fact rather than using within-sample reference measurements. We adopt its affine view of the correction but estimate it per session from references, which is what lets it track fouling drift.
A second group constrains the design. It shows that device heterogeneity is real, that it is correlated with data, and that dropping weak devices hurts. FLHetBench supplies real latency and availability profiles from about 10,000 phones, and under severe heterogeneity reports 35.75% OpenImage accuracy for BiasPrompt+ against 14.84% for FedAvg. WHALE-FL schedules subnetwork widths from system speed and Fisher information, running 1.3–2.1× faster than FedAvg and reaching 89.29% against 87.42% for HeteroFL on MNIST. Adaptive Local Training stops local epochs when local features drift from global ones. The survey on computationally-constrained heterogeneous devices notes that system and data heterogeneity are almost always studied separately. None of these methods asks whether the device being thinned or dropped is the one holding the hardest matrix, and that is the coupling we test.
On the sensing side, online e-nose drift compensation adapts extreme learning machines sequentially over the 36-month UCI gas-drift data. It shows that drift correction can be cheap and online, but it runs on one machine. The electrode papers on exfoliated screen-printed glucose sensors and FeMn@C nanofibers for TBHQ in oil show that materials reduce interference: peak-current changes stay under 10%, and spike recoveries in soybean oil run 99.3–102.6%. They also show the limits of materials alone. The glucose sensor was characterized at pH 13, and no material adapts by itself to a new matrix or to aging. Those recoveries are our reference for what lab-grade accuracy looks like.
Research Program
The work has three stages. The first builds a benchmark that couples matrix harshness with device tier, since the file confirms none exists. Clean response curves are passed through mechanism-based matrix transforms: channel gain suppression, smooth absorbance baselines, slowly growing fouling gain with capacitive offset, and overlapping interferent peaks. Transform severity for each client is drawn jointly with a device profile from the FLHetBench FL-Device database, with a tunable correlation $\rho$ between low hardware tier and high severity. The clean curves come from two sources: the UCI gas-sensor drift data, where the real temporal drift is kept and client-level matrices are layered on top, and voltammetric calibration curves of the kind reported in the TBHQ and glucose papers. Setting $\rho = 0$ recovers the independent-heterogeneity case that existing benchmarks assume.
The second stage builds the method. It consists of the closed-form matrix layer; a small one-dimensional convolutional feature extractor sized so that one forward pass fits microcontroller-class memory; the summary-statistic ridge head; and gradient updates to $\phi$ and $B$ on capable devices only. We will also analyze identifiability. With too few reference measurements, gain and basis coefficients trade off against the analyte signal, so we will measure error as a function of the number of references per session and the rank of $B$.
The third stage is an optional small wet-lab validation. It would use several screen-printed electrodes of different ages across a few food matrices, with spike recovery as the ground truth. This stage settles whether the gain-plus-baseline model is adequate for real matrices, and we treat it as a check on that assumption rather than as the main evidence.
Baselines
The naive reference point is FedAvg with straggler eviction on uncorrected signals. It uses the same feature extractor and head, trained end to end, and drops any client that misses the round deadline set by its FLHetBench profile. Beating it is necessary but proves little.
The strongest existing competitor for heterogeneity is width-sliced federated training in the style of HeteroFL and WHALE-FL. It is the best-documented way to keep weak devices in training, with WHALE-FL's reported 89.29% against 87.42% for HeteroFL on MNIST. It is also the method our coupled-heterogeneity argument targets directly, because it shrinks capacity exactly where matrices are hardest. We will run WHALE-FL's utility-based width scheduling with the same device profiles, the same wall-clock budget, and the same hyperparameter search budget. Devices below the memory needed for even the thinnest subnetwork's backward pass are excluded, which is what that design implies. The strongest heterogeneity-aware aggregation baseline is BiasPrompt+ from FLHetBench, whose reported 35.75% OpenImage accuracy is the best figure in the file for realistic device and state heterogeneity. We will adapt its staleness-aware gradient projection to our regression setting.
The strongest competitor for matrix correction is Federated Reconstruction with a local input-affine layer. It has the same parameter split as our model, but the local part is rebuilt by gradient steps, and devices that cannot backpropagate are dropped. We also run Distributed Cluster ComBat harmonization of features followed by FedAvg, and a centralized model trained on pooled data as an upper bound.
For the gas-drift data specifically, we found no reported federated result on Lacuna. The ODAELM paper reports only that online adaptation stays within a few points of offline retraining, with no exact figure given in the file. We will therefore rerun that method centrally as a single-machine reference.
Two ablations isolate our ingredients. The first keeps the closed-form matrix layer but drops all non-backprop devices, which isolates the value of the summary-statistic head. The second keeps all devices but replaces the physical layer with an unconstrained local linear input layer fitted by least squares, which isolates the value of the physical structure. Our method counts as winning if it lowers concentration RMSE on the worst-severity decile of clients below every baseline at equal wall-clock and communication budget, at $\rho \geq 0.5$, without losing to them on the median client.
Why not just…?
Why not just calibrate each device locally and run ordinary federated learning among the devices that can? This is the most serious threat. It is exactly our first ablation, and if it matches the full method, our heterogeneity argument fails. The risk is that local calibration alone might remove enough of the matrix effect that excluding weak devices costs little. We expect it to fall short at high $\rho$. Correction removes gain and baseline but not non-affine interferent overlap, which the shared head can only learn if it sees those matrices. The sweep over $\rho$ and interferent strength settles this directly. If the gap stays small even at $\rho = 0.8$, we will report that summary-statistic participation is unnecessary and reduce the contribution to the benchmark and the matrix layer.
Why not just use Federated Reconstruction, which already keeps parameters local? Its local rebuild is gradient-based, so the microcontroller tier cannot run it. Its local parameters are also unconstrained, which our identifiability analysis predicts will absorb analyte signal when references are few. The paper reports that a single reconstruction step recovers most performance on MovieLens, and it is not obvious that this carries over to sensing. The comparison at 1, 3, and 10 reference measurements per session shows whether physical structure matters or whether a free local layer is enough.
Why not just thin the model for weak devices, as WHALE-FL does? Width slicing keeps weak devices only if they can afford some backward pass, and it gives them the least capacity where deconvolving overlapping backgrounds needs the most. The worst-decile RMSE comparison, broken down by device tier, tests this. If sliced models match ours on harsh matrices, then capacity was not the bottleneck and the method's case weakens.
Why not just stream raw data to the cloud and train centrally? The centralized run is included as an upper bound, not a straw man. The research file gives privacy and uplink-energy reasons against pooling in wearable and remote settings. Our question is how much of the centralized accuracy a fleet can recover without pooling, and the answer is reported as a fraction of the centralized worst-decile RMSE.
Why not just dilute or use standard addition physically? Dilution pushes trace analytes below the limit of detection, and repeated physical spiking cannot run on continuous wearables. Our method needs only occasional reference measurements, and it inherits standard addition's logic rather than its workflow. The wet-lab stage compares recovery from our model against classical standard addition on the same electrodes.
Proposed Experiments
The decisive experiment sweeps the tier-severity correlation $\rho$ from 0 to 0.8 on the coupled benchmark. The prediction is that the advantage of our method over width slicing and over the calibrate-then-drop ablation is near zero at $\rho = 0$ and grows with $\rho$, concentrated in the worst-severity decile. A flat curve would falsify the central claim.
A second check replaces the physical layer with an unconstrained one while varying the number of reference measurements per session. We predict the physical layer wins most when references are scarce.
A third experiment uses the real temporal drift in the UCI gas data. It tests whether per-session closed-form gain re-estimation tracks fouling-like drift as well as the centralized online extreme learning machine.
Across all runs we report worst-decile and median RMSE, the fraction of the device fleet that contributes to the shared model, per-device peak memory and energy estimated from the FLHetBench profiles, wall-clock time to a target RMSE, and uplink bytes. The summary-statistic uplink is a $d \times d$ Gram matrix plus a length-$d$ vector, where $d$ is the feature dimension, so we will report it against the size of a gradient update. The wet-lab stage reports spike recovery in two oil matrices and compares it with the 95.5–102.6% range the TBHQ sensor achieved under laboratory protocol.
References
- Federated Reconstruction: Partially Local Federated Learning
- Decoupled Low-Rank Adaptation for Robust Federated Fine-Tuning
- Distributed Harmonization: Federated Clustered Batch Effect Adjustment and Generalization
- FLHetBench: Benchmarking Device and State Heterogeneity in Federated Learning
- WHALE-FL: Wireless and Heterogeneity Aware Latency Efficient Federated Learning over Mobile Devices via Adaptive Subnetwork Scheduling
- Adaptive Local Training in Federated Learning
- Federated Learning for Computationally Constrained Heterogeneous Devices: A Survey
- Online Sensor Drift Compensation for E-Nose Systems Using Domain Adaptation and Extreme Learning Machine
- Construction of Highly Active Interfaces on Screen-Printed Carbon Electrodes via Controllable Electrochemical Exfoliation for High-Performance Flexible Enzyme-Free Glucose Sensing
- A Carbon Nanofiber Electrochemical Sensor Made of FeMn@C for the Rapid Detection of Tert-Butyl Hydroquinone in Edible Oil