Machine learning is the practice of fitting parameterised models to data rather than hand-coding rules, and by 2026 it has become the default computational method across materials science, quantum error correction, wireless systems and imaging, with the open questions now about data economics, verification and where the arithmetic physically runs.
Summary
Machine learning (ML) replaces explicitly programmed rules with a parameterised function whose parameters are fitted to data. The family covers supervised learning (fit a labelled input-output map, as in the 1D convolutional network trained to detect RF jamming from IQ samples), unsupervised learning (find structure without labels, as in the PCA plus K-means framework used to characterise die-to-die variability in ferroelectric HfZrO capacitors and predict the behaviour of unseen dies), and large-scale self-supervised pretraining, where a single model is trained on a broad corpus and then adapted to many downstream tasks. WiFo-2, pretrained on 11.6 billion channel state information points, is the wireless instance: it shows scaling-law behaviour, performs zero-shot channel reconstruction better than fully supervised task-specific models, and reaches state-of-the-art results on nine tasks using 1% of the training samples supervised models need.
Three things decide whether an ML model works on a given problem. The first is the hypothesis class and its inductive bias: neural networks, kernel machines and Gaussian processes each encode different assumptions, and injecting known physics sharpens them. A tensorial kernel support vector machine, an interpretable unsupervised method, identified string order in symmetry-protected topological phases directly from trapped-ion experimental data; symmetry-equivariant quantum circuit ansaetze built by twirling encode finite-group and Lie-group symmetries of a PDE as a bias; and quantum Gaussian processes use a physics-informed prior over unknown quantum transformations.
The second is data. Training sets are expensive to produce, and the emerging finding is that much of the cost is wasted: electronic structure datasets generated by Kohn-Sham DFT contain large redundancies attributable to low intrinsic dimensionality, so even random pruning barely degrades accuracy, and coverage-based pruning preserves chemical accuracy while using up to two orders of magnitude less data and cutting training time by a factor of three or more. Where raw data cannot be pooled at all, federated learning trains a shared model across devices without moving the data.
The third is where the arithmetic runs. Digital electronics is the incumbent, but the same mathematics has been mapped onto other physical substrates: passive diffractive layers that compute at the speed of light with no power beyond the illumination, multiple scattering through modulation planes that synthesises both linear and nonlinear transforms at milliwatt continuous-wave power, InP photonic integrated neurons projected at 47 TMAC/s and 12 pJ/MAC, and thermodynamic hardware that treats noise as a computational resource for diffusion models, Bayesian networks and Monte Carlo sampling ref. Substrate choice does not change what ML is; it changes the energy and latency at which it can be deployed.
Evidence base
- 2026-05-05: WiFo-2, pretrained on 11.6 billion CSI points, shows scaling-law behaviour, beats fully supervised task-specific models at zero-shot channel reconstruction, and reaches state of the art on nine wireless tasks with 1% of the supervised training samples, with a functional hardware prototype.
- 2026-05-04: Coverage-based pruning of electronic structure datasets preserves chemical accuracy and generalisability while using up to two orders of magnitude less data and cutting training time by a factor of three or more, explained by the low intrinsic dimensionality of the data.
- 2026-05-05: A federated 1DCNN trained on over-the-air IQ samples achieves 97% accuracy and 97% F1 for 5G RF jamming detection, outperforming centralised MLP, 1DCNN, SVM and logistic regression baselines without sharing raw data.
- 2026-05-04: In quantum error decoding, neural network decoders are grouped with clustering, matching and belief propagation as inadequate, and a randomised belief-propagation variant delivers a 2 to 8 order-of-magnitude accuracy gain for topological codes.
- 2022-01-13: A simulated 64:64:10 all-optical InP neural network reaches 89.5% best-case MNIST accuracy at 10 GS/s, with 47 TMAC/s (about 2.5x a then-state-of-the-art GPU), 12 pJ/MAC (about 2x better) and full-system energy below 20 pJ/op above 18 synapses per neuron.
- 2026-05-04: Quantum machine learning is described by its own practitioners as ‘in many ways stuck’, with quantum Gaussian processes proposed to supply a simple, interpretable and scalable framework for learning on quantum data.
Open questions
- Does the low intrinsic dimensionality finding generalise beyond electronic structure data, or are the two-orders-of-magnitude pruning ratios specific to DFT-generated datasets?
- Does the scaling-law behaviour reported for a wireless foundation model hold when the pretraining corpus is broadened beyond 11.6 billion CSI points, and does the zero-shot advantage survive on hardware in the field rather than on a prototype?
- Can any alternative substrate move from simulated multi-layer networks and MNIST-scale tasks to a fabricated system that holds the claimed 12 pJ/MAC and 47 TMAC/s at useful model sizes?
- Will certified-training and mutation-testing methods mature fast enough to let learned components enter safety-critical or regulated loops, or will hand-verified classical algorithms keep winning those slots as they did in topological-code decoding?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.