An electrical power grid could serve as a fixed computational step inside an image-classification system, according to a simulation preprint. In the tested pipeline, digitally encoded voltage perturbations were injected at controllable distributed-energy-resource nodes, propagated through the fixed grid operator, and read out as current responses at node 8. A linear decoder then turned those responses into class probabilities.
The strongest result came from a limited test. On the first 400 test samples from each main dataset, the PLECS time-domain pipeline reached 91.50% accuracy on MNIST and 82.25% on Fashion-MNIST. An ideal surrogate reached 95.25% on MNIST. The paper reported no confidence intervals for these results.
The grid in the pipeline
The pipeline treated the network as a fixed operator between an encoder and a linear decoder. The benchmark contained 14 buses, 17 transmission lines and 3 power transformers. Converter-interfaced distributed energy resources were placed at nodes 1, 2, 3, 6 and 8; four nodes were controllable inputs, and node 8 was the principal read-out.
That design produced a notable representation difference, but not a universal hardware claim. The closest fully connected replacement had 30 hidden units and used 12,340 task-trained parameters. The physical operator used 933 surrogate parameters, a reported 13.2-fold reduction in the core-operator representation. The comparison does not establish universal parameter, hardware, energy or latency advantages, and the surrogate parameters were not optimized for MNIST classification.
To build the surrogate, the study used 1,000 independently generated PLECS input sequences. Accuracy checks used confusion matrices, which count predicted and ground-truth label combinations, and calculated per-class accuracy from the proportion correct within each ground-truth class.
Performance depended on the setup
The benefit also depended on what the grid was asked to process. The physical pathway scored higher than the static pass for MNIST and Fashion-MNIST. In the CIFAR-10 evaluation, stronger upstream feature extraction was associated with better encoded separability and a smaller physical-versus-static difference. The result points to a representation-dependent role for the grid.
The study also used a Butterflies-and-Moths dataset containing 13,094 training images and 500 test images across 100 species. But the evaluation consisted of 25 independent experiments, with 10 species selected at random for each experiment. It therefore was not a 100-class evaluation.
Stress tests showed that robustness depended on the particular network change. A 10% increase in line-7 impedance left mean accuracy at 0.922, while removing load 6 left it at the 0.915 benchmark. Removing line 10, however, produced mean accuracy of 0.495. These results describe a finite tested operating envelope, not every possible grid condition.
Measurement noise created another boundary. With additive white Gaussian noise on the sampled current responses, mean accuracy ranged from 0.917 to 0.910 as the noise standard deviation moved from 0.001 to 0.05 per unit, the study's normalized scale. At 0.1 per unit, accuracy was 0.810. Only the reported noise conditions were evaluated.
The number of active encoding channels mattered too. Reducing the setup from four channels to three produced mean accuracy of 0.910, close to the four-channel benchmark of 0.915, while two channels produced 0.802. Two tested kernel swaps produced 0.917 and 0.910. The study describes this as a configuration sensitivity analysis within one benchmark system.
Sharing the operator across tasks was more uneven. Spatial multi-tasking produced mean per-task accuracies of 0.784 and 0.851 after five training epochs, rising to 0.858 and 0.881 after 20 epochs. Temporal multi-tasking produced 0.965 and 0.938 after five epochs, and 0.962 and 0.941 after 20. In the study's surrogate evaluation, temporal reuse gave each stream access to the full operator, while spatial sharing split access across streams. The results do not establish hardware throughput, crosstalk, latency or energy savings.
The next test is outside the model
The work is an arXiv v1 preprint dated 6 October 2026, and all reported results came from the IEEE 14-bus benchmark and time-domain simulation. No physical grid deployment or laboratory-scale validation was reported. Laboratory testing with programmable converters was identified as a next step.
That leaves the central practical questions unanswered. The 13.2-fold figure concerns the representation of the core operator, not a universal hardware, energy or latency advantage. The multiplexing results likewise do not establish throughput or deployment performance. The evidence supports a narrower conclusion: the tested grid operator added task-relevant transformations in simulated image-classification pipelines, but did not establish general superiority.
The next step is to see whether programmable converters can reproduce these results outside the simulated IEEE 14-bus setting. Until then, the study is best read as a simulation proof of concept for using electrical infrastructure as a constrained physical operator, with its usefulness depending on the input representation and the way the network is configured.
Paper data and sources
Original title: Infrastructure-Native Computing with Electric Power Grids
Authors: Yubo Song, Subham Sahoo, Freja Basse
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-10-06
DOI: Not available
Original paper · Full text