Broadband encoding and high-speed probabilistic bit generation with integrated microwave neurons Authors: Bala Govind, Pablo Raigoza, Alyssa Apsel Published: 2026 ABSTRACT AbstractIntelligence at high data rates depends on extracting meaning as information arrives. Although radio signals are analog, most systems digitize them before feature extraction, adding latency and power overhead while limiting processing speed to digital clock rates. Instead, we find that incoming data can itself reprogram coupling between frequency modes in an ensemble of microwave waveguides, encoding spectral features across several gigahertz. This milliwatt-scale Microwave Neural Network (MNN), manufactured in Complementary-Metal-Oxide-Semiconductor (CMOS) technology, expresses data tokens across a spread spectrum, making relationships between successive tokens easier to detect and enabling faster downstream decisions with lightweight digital processing. We also find that gigabit-per-second bitstreams, like 8-bit pixel data, drive the MNN to produce probabilistic bits whose bias reflects the input pattern. This analog dithering preserves detail in compressed data, potentially enabling richer satellite imaging at one-eighth the bit-rate under bandwidth constraints.Subject terms: Electrical and electronic engineering, Computational science, Computational nanotechnology, Electronic and spintronic devices Introduction The ability to efficiently encode small amounts of information by forming complex relationships among them and later decode these relationships is fundamental to fast communication systems1,2, cryptography3,4, and artificial intelligence5,6. However, this type of processing at ever-higher bit speeds quickly saturates the memory bandwidth7 of computational accelerators like graphics processing units (GPUs)8, tensor processing units (TPUs)9, and neural processing units (NPUs)10. In all of these platforms, data is represented as digital bit-strings that must be handled by logic circuits whose speed is limited by a clock. On top of these hardware constraints, modern machine-learning architectures capture relationships between data tokens by embedding them into high-dimensional vector spaces and performing large matrix operations to infer context. These include Transformer-based models11 and recurrent neural networks such as Long Short-Term Memory units12. While they are effective at modeling long-range dependencies and contextual relationships, as data rates push into the multi-gigabit-per-second regime, both latency and power consumption increase steeply due to the combined costs of data movement, memory access, and clocked matrix operations13. Moreover, real-world signals are analog. So far, it has been necessary for sensor outputs to be sampled, digitized, and quantized before any inference can occur. This requires high-bandwidth analog-to-digital converters (ADCs), which become increasingly costly in power and area14 on a semiconductor chip. To overcome this, intelligence must move into the analog front end, extracting information before full digitization while signals still retain their rich broadband features. While several non-digital approaches have been explored for feature extraction in time-series and image data—including optical matrix multipliers15,16, memristor crossbar arrays17,18, thermodynamic computers19 and quantum annealing processors20—these systems face practical challenges. They typically require substantial external high-speed readout circuitry, making monolithic integration of sensing and processing on a single complementary-metal-oxide-semiconductor (CMOS) chip difficult. Notably, their computational tasks have largely been limited to low-data-rate tasks like combinatorial optimization, image classification and voice recognition. We recently introduced the microwave neural network (MNN)21, a CMOS computing platform that bypasses sequential, clocked time-domain processing. It instead performs computation through nonlinear interactions among microwave frequency modes spanning several tens of gigahertz. In our initial demonstrations, slow 1/0 parametric control bits dynamically altered coupling between waveguides, enabling the MNN to respond differently to high-speed broadband data streams and wireless analog signals and reconfigure its behavior for distinct tasks without new fabrication. In particular, the MNN was parametrically configured for RF modulation classification and radar target detection, where hard-on/hard-off parametric coupling sequences applied to the control port were co-optimized with a lightweight backend. We further hypothesized that only a few gigahertz of readout bandwidth would be sufficient for machine-learning inference, rather than the full microwave spectrum of the MNN’s output. However, the physical mechanisms by which information interacts with the MNN’s oscillatory modes and becomes encoded within its broadband spectral response remained unclear. Understanding these mechanisms motivates the present study. Here, we show that nonlinear mode coupling in the MNN enables two distinct mechanisms through which broadband signals can be encoded and interpreted (Fig. 1). In the first, analog data tokens encoded as pulse trains control coupling between waveguides through their amplitude, modulation frequency, and duration. Although these signals occupy only a few hundred megahertz of bandwidth, they instantaneously reshape the MNN’s broadband spectral response across several gigahertz. This exposes features of the input and relationships between successive tokens that can be interpreted by lightweight embedding and inference models for edge decision-making. In the second, the waveguides are driven directly by gigabit-per-second bitstreams without parametric modulation. In this regime, broadband input features are redistributed across the MNN’s output spectrum, enabling recovery of information from a narrow output band via low-bandwidth analog-to-digital conversion. We further find that repeated injection of the same bitstream produces input-dependent probabilistic responses arising from phase-sensitive interactions between the incoming data and the MNN’s free-running oscillations. This native hardware dithering22,23 generates probabilistic bits whose statistics depend on the input pattern and suggests a route toward transmitting high-resolution satellite imagery using only a fraction of the conventional communication bandwidth. Two types of feature extraction with an integrated microwave neural network (MNN). Microwave waveguides in CMOS are coupled to produce a neural network of oscillatory frequency modes. Waveguides B, C and D respond linearly to input signals, while A operates nonlinearly. The interacting oscillatory modes are sustained by a saturable gain medium and can be reprogrammed to extract features of incoming signals in two ways. a Analog data tokens, with a bandwidth of a few hundred megahertz can be used to control switches that modulate the transient coupling between pairs of waveguides. Parametric upconversion broadens the MNN’s output comb across several gigahertz. This process embeds distinguishing features of the incoming data tokens into the spectral response, enabling machine-learning inference to identify relationships between tokens. b Instead, broadband data of several gigabits/second could also be directly fed into the waveguides. MNN distributes features from across the input spectrum throughout its instantaneous comb-like output, so each narrow spectral band contains information from many input bands. Consequently, only a low-speed readout bandwidth of a few hundred megahertz is required to capture the essential features of the broadband input. CMOS complementary metal-oxide semiconductor, DC 0 Hertz, GHz gigahertz. Results Parametrically-coupled oscillations encode relationships between data tokens We find that when an incoming pulse pattern modulates the coupling between the MNN’s waveguides in time, it drives a parametric mixing process that extracts compressed spectral features of the input (Fig. 1a). We study this behavior by computing spectrograms of the MNN’s oscillations measured at any of the MNN’s waveguides’ output ports (Fig. 2a). Transformation of the MNN’s time- and frequency-domain response as its output comb is reshaped by radio-frequency pulse parameters. a Radio-frequency pulse sequences are applied to the MNN's token port, tuning the coupling between its linear and nonlinear waveguides. These inputs reshape the instantaneous structure of the MNN's output comb, which lies between approximately 10 and 14 GHz, and redistribute power within narrower frequency bands. The output is mixed down for observation within a lower readout bandwidth, and its time-varying frequency content is examined using a spectrogram. The transition between the responses from Token 1 and Token 2 occurs almost immediately, indicating near-instantaneous nonlinear behavior in the MNN. Each token’s distinct broadband response is observed from DC to 2.5 GHz. b The time-averaged steady-state output comb varies in response to all three parameters of the RF pulse tokens. The pulse amplitude, βTok, has the strongest influence on mode-coupling strength and comb bandwidth (b.i), followed by the duration, TTok, of the pulse train (b.ii), while changes in the frequency, FTok, of the token exert a slightly weaker effect (b.iii). GHz Gigahertz, MHz Megahertz, ns nanosecond, dBm decibel-milliwatt. In this configuration (experimental setup described in “Methods” section “Chip design, experimental data acquisition and baseline generation” and Supplementary Fig. 9), radio-frequency pulse trains are sequentially applied to the parametric switching port as analog “data tokens”, with amplitude (0–1 V), duration (20–100 ns), and pulse frequency (20–100 MHz) varied. The spectrograms of the raw tokens themselves are largely uninformative, as they exhibit structure primarily below 200 MHz. In contrast, when the same tokens modulate the MNN’s parametric coupling—altering interactions between linear and nonlinear modes—their features instantly unfold into a broadband frequency comb at the output. In our circuit, we find that the most expressive portion of this comb lies between 10 and 13 GHz, well beyond the bandwidth of conventional analog-to-digital converters (ADCs). To access this information, we down-convert the comb to baseband (0–3 GHz) using a passive mixer and local oscillator, enabling time-domain capture using modest-bandwidth readout (in this chip version, via an oscilloscope). As Fig. 2a shows, each token elicits an immediate response through input-driven parametric nonlinear coupling. The spectral reshaping is transient, and the individual pulse parameters have a measurable influence on the comb’s properties. By repeatedly injecting individual tokens and recording the MNN's steady-state spectral output, we determine how each pulse parameter influences the resulting bandwidth (Fig. 2b). The amplitude (βTok) has the strongest effect (Fig. 2b.i), as it directly scales the coupling strength between waveguides. Small amplitudes (βTok < 0.5 V) yield weak mode interaction and a narrow comb, whereas amplitudes approaching 1 V produce dense intermodulation products spanning several gigahertz. Similar bandwidth expansion occurs when either the pulse frequency (FTok) increases (Fig. 2b.ii), or the pulse duration (TTok) lengthens (Fig. 2b.iii). These trends are consistent with simulations using the coupled-mode model discussed in the “Methods” section, “Modeling feature extraction mechanisms through the MNN’s nonlinear dynamics,” and Supplementary Fig. 8. The central question is how to exploit the MNN’s sensitivity to pulse parameters to encode meaningful information. We address this by measuring how reliably the spectral response generated by one token supports the identification of another. Specifically, for each ordered token pair (a, b), we define an association score as the fraction of cross-validation folds24 in which a linear classifier correctly identifies token a from the MNN response to token b. This score reveals latent relationships between successive tokens and can be used to estimate likely continuations in a sequence. We illustrate this concept through a turn-based navigation game we developed (Fig. 3a). In this game, two submarines alternately reflect sonar signals that bounce off boundaries, each attempting to intercept the returning beam. A successful pair of returns defines a rally, and the sequence continues until a submarine fails to intercept the signal. Each interception or miss constitutes an event in a temporal sequence, analogous to a token stream. The submarine’s position (x, y) and reception angle (θ) define an instantaneous state in a three-parameter space (x, y, θ) (Fig. 3b.i). Some state transitions are more compatible than others due to geometric constraints and finite movement speed. By simulating numerous optimal-play scenarios (“Methods” section “State Compatibility Matrix and mechanics of the submarine navigation game”), we derive a compatibility matrix that encodes the relationships between states as embedding vectors (Fig. 3c.i). MNN maps navigation instructions to radio-frequency pulse trains. a A turn-based navigation game in which submarines alternately intercept and reflect sonar beams that arrive after bouncing off the sea surface, surrounding walls, and the other submarine. Each successful pair of interceptions constitutes a rally that continues the game, while a missed interception ends the game. b There exist two 3D coordinate state spaces that can be mapped to one another: one defines a submarine’s instantaneous position, specified by its x--y position and the angle (θA or θB) at which the sonar signal reflects off it (b.i), and a corresponding microwave pulse-train vocabulary understood by the MNN + linear layer pipeline, defined by amplitude (βTok), pulse frequency (FTok), and duration (TTok) (b.ii). c In the sequence of moves that constitute perfect play, each submarine, knowing the dynamics of subsea wave reflections, moves to optimal positions to intercept transmissions reflected from its counterpart (c.i). Submarine states in each sequence, like data tokens, have embeddings expressed through a move compatibility matrix. c.ii The MNN responds to low-speed RF pulse trains, and its expansive nonlinearity produces a several-gigahertz bandwidth expansion that reveals the features of incoming tokens. A linear layer is applied to the spectrograms of the MNN's responses to infer relationships between tokens based on spectral similarity and produce a token association matrix. c.iii A greedy search assigns descriptions for submarine states to a subset of pulse tokens, ignoring those with lower embedding similarity. These mappings allow a submarine guided by the MNN to sustain multiple rallies without missing transmissions and without requiring explicit knowledge of the underlying physics of the environment. The MNN’s state is similarly defined by three pulse-excitation parameters (βTok, FTok, TTok) (Fig. 3b.ii). To establish relationships between these states, we analyze pairwise MNN responses to pulse tokens, each producing unique spectral signatures across a 2.5 GHz bandwidth. Feeding these spectrogram features into a trained linear layer reveals how likely one token is to follow another, yielding the Association Matrix in Fig. 3c.ii (see “Methods” section “Generation of the token Association Matrix relating microwave pulse-train tokens” for its derivation). The resulting embedding is sparse and structured, showing clear affinities between specific MNN states, and weak relationships between others. In contrast, the Association Matrix constructed from raw token spectrograms used as the baseline for comparison, which contains only sub-200 MHz content, is blurred and reveals indistinct relationships between tokens (see Supplementary Fig. 10). We then test whether a submarine’s transient state can be expressed within the MNN’s pulse-parameter space. We perform a Greedy Search25 to align the submarine Compatibility Matrix with the MNN Association Matrix. The algorithm selects the pulse tokens that best reproduce allowed submarine transitions, excluding those that do not capture valid state progressions (Fig. 3c.iii). A detailed formulation of this assignment process is provided in Supplementary Section 1 and Supplementary Fig. 1. Finally, we allow the MNN to guide one of the submarines (submarine B) to evaluate whether our inferred state-to-token mappings enable successful navigation (Fig. 4a). Submarine A acts as a perfect opponent. The MNN-controlled submarine processes tokens representing both submarines’ current states, generates spectrogram features, and the backend classifier chooses the next move. The exchange continues until interception fails. Performance is quantified by the number of rallies survived (Fig. 4b). We test six motion speeds normalized to distance traveled along the grid per rally. With a limited embedding vocabulary of only 25 states, excluding reflection angle, the MNN performs only slightly above baseline (Fig. 4b.i). Expanding to 75 states by incorporating angle information increases survival relative to the baseline (Fig. 4b.ii) by up to 33% at the highest speed (25 rallies with the MNN versus 18 with the baseline). Higher submarine speeds result in longer survival, as expected. MNN-guided decision-making for a submarine navigation task. a Radio-frequency pulse-train tokens replace traditional direction-pad commands, enabling the MNN + linear layer to interpret an opposing submarine’s state and respond intelligently. The MNN submarine (with its state encoded as a pulse token) bounces a signal off into the environment and then receives the state of submarine A (also encoded as a pulse token) that intercepts it, . Both tokens are sequentially fed to the MNN. Through its expansive nonlinearity, and using a trained linear layer, it infers where to go next to intercept the signal reflected from submarine A-allowing the rallies to continue. b The performance of MNN-assisted inference is compared with a baseline without the MNN. Survival is observed across a range of possible travel speeds for the MNN-guided submarine. b.i When the token vocabulary encodes only x--y coordinates with 25 RF pulse-train tokens, the MNN offers little advantage over the baseline. b.ii When the vocabulary is expanded to also encode the reflection angle (θ), representing a submarine’s full state (x, y, θ) with 75 tokens, the MNN-guided submarine survives 33% more rallies than the baseline. These results demonstrate that the MNN can extract feature-rich representations from sequential inputs, and that the inferred state-token mappings are sufficient to support predictive decision-making in a dynamic setting when interpreted by a lightweight backend decoder. Performance improves systematically as additional state information is incorporated into the embedding space, indicating that the MNN’s spectral responses encode meaningful relationships between successive states. In a separate demonstration, demonstrated in Supplementary Section 2 and Supplementary Figs. 2–3, we show that the MNN can also infer linguistic structure directly from messages encoded in analog pulse tokens. By mapping a small vocabulary of tokens to Parts of Speech (PoS) like nouns and verbs, the MNN can be used to identify the PoS of successive tokens from their combined spectral response and to predict the next valid PoS within common English sentence templates. Together, these capabilities define a scheme where structured information is encoded in broadband spectral features, which a lightweight backend uses for token matching and next-token prediction. Generation of fast probabilistic bits Beyond developing embedding vectors between megahertz-rate pulse trains, the MNN also supports a fundamentally different mode of feature extraction that helps interpret high-speed communication signals using low-bandwidth readout. Here, we inject multi-gigabit/s bitstreams directly into a waveguide, as shown Fig. 1b. Any pair of ports connected to the waveguides can serve as the input and output for this operation, as shown in Fig. 5a (and in greater detail in Supplementary Fig. 9b). This ultrafast excitation produces distinct, instantaneous, feature-rich time-domain responses and a complex comb-like output spectrum. A key objective is to determine whether useful broadband features can be redistributed across the multi-gigahertz output spectrum into a compressed spectral representation, such that only a narrow portion of the MNN’s spectrum may need to be down-converted to baseband for probabilistic reconstruction. MNN generates low-speed probabilistic bits from high-speed bitstreams. a Broadband data is fed into one on-chip waveguide (D) and read out from another (A). The MNN redistributes broadband input information across its comb-like output spectrum, such that each narrow sub-band contains representative features of the full input. As a result, a single narrow band can be sensed using a low-bandwidth baseband receiver. b An 8-bit symbol at 2.5 Gbps is injected into a waveguide, and the response is down-converted through a passive mixer that selects a narrow sub-band. A 1-bit quantizer operating at one-quarter of the bit rate captures two samples, assigning a 0 or 1 based on signal polarity. Across repeated measurements, these samples fall into one of four possible 2-bit outcomes (10, 01, 00, 11), grouped as Dynamic (10, 01) or Static (00, 11) states. This compresses each 8-bit input into a single probabilistic bit (p-bit) defined by the relative likelihood of Static versus Dynamic outcomes (Static Ratio). c Input-phase-dependent responses in the MNN give rise to p-bits. Small differences in input arrival phase relative to the oscillation alter the transient baseband response, yielding the four possible 2-bit outcomes after sampling and quantization. d Repeating each input many times yields a probability distribution over outcomes for every 8-bit pattern. Sorting all 256 patterns by Static Ratio defines an MNN-encoded grayscale, where a higher Static Ratio yields more dark pixels and a lower Static Ratio yields fewer dark pixels. Gbps Gigabit per second, Mbps Megabit per second. We deliberately use a sampling chain that operates far below the Nyquist rate (Fig. 5b). A 1-bit quantizer captures only two samples from the MNN output during each 8-bit symbol interval, effectively one observed bit for every four data bits. For example, at 2.5 Gb/s, an 8-bit pattern spans only 3.2 ns, so a baseband ADC operating at 625 MS/s (one-quarter rate) captures merely two samples in that interval. Limited readout bandwidth means only a few samples are acquired during the short excitation window in which the MNN’s nonlinear dynamics unfold. Under these conditions, we observe that the same 8-bit input can produce different 2-bit outcomes (00, 01, 10, 11) across repeated trials. This variability originates not from environmental noise and temperature fluctuations (Supplementary Information sections 3 and 4 and Supplementary Figs. 5 and 6) but from the MNN’s inherent phase sensitivity. The transient output waveform depends on the precise timing at which external pulses arrive relative to the oscillation cycle (Fig. 5c). An immediate consequence is that the same 8-bit bitstream, when injected into the free-running MNN, produces a different output waveform each time it is down-converted to baseband. In conventional CMOS oscillators, impulse sensitivity functions26 (ISFs) indicate which points in the oscillation cycle are most vulnerable to disturbance, linking external interference to timing jitter and injection locking. In the MNN, however, the presence of multiple strongly coupled nonlinear modes makes such an ISF-based description intractable. Instead, a generalized coupled-mode formulation explains why the RF oscillation being much faster than the data rate makes small timing shifts in symbol arrival naturally lead to distinct transient baseband responses (Fig. 5c). Surprisingly, despite this extreme undersampling, the outcomes are not random. Performing an experiment where we feed each input pattern into the MNN hundreds of times reveals a unique probability distribution over the four possible states. Each 8-bit pattern produces a biased mixture of Dynamic outcomes (10 or 01) and Static outcomes (00 or 11). Thus, the MNN maps a high-speed 8-bit symbol to a single probabilistic bit (p-bit27,28) whose skew toward Static or Dynamic states still carries meaningful information. Importantly, unlike the first demonstration involving token relationships, no spectrogram extraction or regression-based backend classifier is required here. We define the Static Ratio as the fraction of trials yielding a Static outcome, providing a characteristic probabilistic signature for each pattern (Fig. 5d). Repeating this experiment for all 256 eight-bit symbols yields an MNN-encoded grayscale representation. Symbols with higher Static Ratios generate a larger fraction of dark pixel outcomes, causing regions dominated by those symbols to appear darker; symbols with lower Static Ratios produce fewer dark pixels, making those regions appear lighter. The corresponding two-sample baseband waveforms obtained by downconverting the MNN response around 10.4 GHz—300 repetitions for each of the 256 patterns—are shown in Supplementary Figs. 13 to 24. Finally, we highlight a promising future application that this p-bit generation could unlock once fully integrated sampling, synchronization and baseband circuits are fabricated. As Fig 6a.i shows that darker pixels are mapped to bit-patterns that make the MNN more likely to produce a Static state (p-bit = 1), while lighter pixels are mapped to patterns that yield a lower Static-state probability (p-bit = 0). These bit-patterns would be fed sequentially into an MNN integrated onboard a satellite, immediately following the image sensor array and encoder. The MNN’s RF response during each input window would be down-converted, and 1-bit quantized in real time, producing one p-bit per pixel. When arranged into a matrix, these 0/1 p-bits form a dithering pattern that appears binary when viewed up close (Fig. 6a.iii). From a distance, these probabilistic pixels blend into a smooth grayscale that preserves key features of the original high-resolution image, like the swirling structure of a storm (Fig. 6a.ii). This dithering scheme arises from the controlled randomness generated by the analog front-end’s response to properties such as sparsity (the distribution of ones in the incoming pattern). It provides a continuum of stable probabilistic states that enable smoother intermediate shading without much post-processing (Fig. 6b.i). This is also similar to the concept of non-uniform sampling analog-to-digital conversion29,30, which takes samples at irregular time intervals to compress data in high-speed systems such as radar. In contrast, the digital dithering baseline produces states that collapse to a few discrete values (Fig. 6b.iii). As a result, only a small number of intermediate gray shades are represented, making it difficult to resolve fine details such as the storm’s eye. Dithering with the microwave neural network (MNN) enables higher-fidelity grayscale using probabilistic bit outputs. a In the proposed signal chain, each pixel’s grayscale value is assigned an 8-bit pattern whose Static Ratio corresponds to the desired shade (a.i). These symbols are then streamed sequentially into the MNN at a rate of 2.5 Gbps. The output is down-converted and quantized to a one probabilistic bit (p-bit) per pixel, producing a dithered image. The spatial density of black versus white output pixels encodes perceived brightness, allowing a smooth grayscale appearance when viewed at a distance (a.iii), capturing the main features of the original image (a.ii). b Comparison with a digital dithering baseline. The MNN produces a continuum of stable Static Ratio outcomes for different bit patterns (b.i), enabling many distinguishable gray levels. In contrast, the digital method collapses all patterns into only a few discrete Static Ratio outcomes, limiting shading resolution (b.iii). Reconstructions show that the MNN preserves fine details such as the hurricane’s eye (b.ii), whereas the digital dithering baseline produces visibly degraded reconstruction through the subsampling signal chain (b.iv). Gbps Gigabit per second, Mbps Megabit per second. While demonstrating this concept in Fig. 6, in reconstructing an image, we assigned each pixel’s p-bit by randomly sampling from a cache of 300 previously measured outcomes of the MNN’s response to the corresponding 8-bit pattern (see “Methods” section “Training and closed-loop inference for the turn-based submarine navigation game”). In actual deployment, however, the MNN could generate these p-bits live, with randomness that naturally follows the same probability distribution. Nevertheless, this approach enables the transmission of just one p-bit per 8-bit symbol, providing an 8 × reduction in image data and could help overcome downlink issues in transmission to bandwidth-constrained edge devices. This does not imply faithful preservation of arbitrary broadband input information within a narrow sub-band; rather, the amount of information that can be retained depends on the spectral density of input features, readout bandwidth, and sampling rate, which together determine the achievable frequency-mode resolution at the output. The choice of local oscillator (LO), which selects the portion of the MNN spectrum down-converted to baseband, is also important (Supplementary Fig. 11a). As shown in Supplementary Fig. 11b, sweeping the LO reveals regions with substantially different reconstruction quality. This behavior is reflected in image-quality metrics Supplementary Fig. 11c). Peak signal-to-noise ratio (PSNR) captures pixel-wise fidelity and correctly highlights the optimal LO band. Spectral and gradient-magnitude correlations better reflect perceptual quality because they are sensitive to edges and texture. Structural similarity index (SSIM) undervalues quality in this context due to the high-frequency dithering noise produced by analog p-bit encoding. In Supplementary Fig. 12a.i, we show how reconstructed image quality improves when multiple passes of the same pixels’ bitstreams are injected into the MNN. The Peak Signal-to-Noise Ratio (PSNR) rises as additional dithered MNN outputs are accumulated. Across all pass counts, it remains consistently higher than the digital dithering baseline (Supplementary Fig. 12a.ii). At each pixel location, the p-bit outcomes are averaged across reconstructions, similar to how geostationary satellites accumulate multiple frames over time. We see the expected N (N is the number of passes) improvement from Monte-Carlo denoising, where multiple stochastic estimates collectively suppress noise and reveal the underlying image structure (Supplementary Fig. 12b). We also extend this approach to color imagery (Supplementary Fig. 12c), where an RGB satellite image of an urban area is reconstructed by independently processing the Red, Green, and Blue channels. A single pass produces analog-TV-like Gaussian noise. However, repeated transmission (for example, 100 passes at 2.5 Gb/s from a geostationary satellite) could enable a high-fidelity reconstruction, revealing tiny features on rooftops and vehicles. Discussion In this paper, we have shown that high-level information, such as data tokens and directional instructions, can be transcribed into short microwave pulses that can be interpreted through the oscillatory dynamics of a microwave neural network. The high-bandwidth MNN and a lightweight linear readout backend digital neural network act as an encoder-decoder pair that maps relationships between sequential chunks of time-series data. This suggests a possible communication scheme in which one MNN could encode snippets of language or navigation instructions into physical pulse tokens, while another MNN could rapidly extract meaning from incoming spectrograms. Although not a cryptographic primitive in the formal sense, such an approach could offer a hardware-intrinsic mechanism for obfuscation: the symbol-to-spectrum mapping may be public, but the physical MNN remains the unique device-specific key needed to recover the message. Our experiments also reveal that fast bitstreams can excite MNN dynamics to produce probabilistic bits directly, indicating the beginnings of a probabilistic generative capability in the MNN. This suggests that the platform could be further optimized for image-processing tasks as an event camera, compressive sensor, or wideband spectrum-monitoring front end. Further reductions in power consumption are possible (see Supplementary Table 4). While this demonstration of the MNN front-end prototype relied on an off-chip decoder (“Methods” section “Chip design, experimental data acquisition and baseline generation”), our ongoing engineering developments focus on integrating passive mixers, variable-gain amplifiers, and a sub-25 mW, 500 MS/s analog-to-digital converter for compressed readout on the same CMOS die. The same platform would also integrate a 500 MS/s digital-to-analog converter for token generation and parametric control of the MNN feature encoder. A lightweight RISC-V microcontroller for low-power frequency-domain inference and training would complete the encoder–decoder pair with sub-1 W total power consumption. This would enable closed-loop supervised training (Supplementary Section 5 and Supplementary Fig. 7) on the same die or co-packaged in a multi-chip module. An instruction-set architecture would coordinate parametric control, signal acquisition, mode-coupling estimation, and backend training and inference. Ultimately, integrated microwave neural networks would help edge systems infer relationships embedded in broadband data with lightweight downstream processing. Such an approach could bring real-time capabilities approaching those of billion-parameter neural networks directly into wideband radio transceivers installed on drones and small satellites, aiming to bypass the speed limitations and power overhead imposed by clocked, sequential operations and data conversion in digital systems (Supplementary Table 5). Methods Chip design, experimental data acquisition, and baseline generation The CMOS Microwave Neural Network platform was fabricated in GlobalFoundries’ 45 nm RFSOI Silicon-on-Insulator semiconductor process. The circuit was designed using the Cadence Virtuoso environment and with the 2.5D EMX tool to model mm-wave performance of on-chip waveguides and couplers. Parasitic resistances and capacitances were extracted using Siemens’ Caliber tool. MNN ver1.0 consists of one nonlinear waveguide (A) and three linear waveguides (B-D). Gigahertz-speed microwave signals are injected through GSGSG interfaces and distributed by miniature quadrature hybrid couplers, allowing reflected signals from the waveguides to interfere and be read out. The network’s computational behavior arises primarily from waveguide A, which contains cascaded nonlinear resonators formed from inductive segments and voltage-dependent nonlinear capacitors implemented with antiparallel diodes. These nonlinear elements make the supported frequency modes highly sensitive to the amplitude and phase of incoming signals. In contrast, the linear waveguides act as reconfigurable transmission lines whose effective lengths can be adjusted by switches. Additional NMOS switches between waveguides enable time-varying (parametric) coupling, dynamically altering interactions among frequency modes. To compensate for loss and sustain strong nonlinear oscillations at tens of gigahertz, cross-coupled NMOS power-amplifier class transistor pairs provide saturable gain. Supplementary Fig. 9 shows the experimental setups used to evaluate both mechanisms of feature extraction with the MNN. Wirebonds provide connections for all RF inputs and outputs, token inputs, power supplies, and the on-chip serial-to-parallel interface (SPI) that initializes the MNN’s state for each experiment. The chip is mounted in a recessed pocket on a compact custom printed circuit board to minimize wirebond lengths. Ground-signal-ground (GSG) and ground-signal-ground-signal-ground (GSGSG) transmission lines are used for RF inputs and outputs to ensure controlled impedance and low insertion loss. The MNN’s oscillator core is biased at 0.9 V. The logic buffers that drive the switches controlling the length of the tunable linear waveguides are biased at 1.6 V, with logic levels of 1/0 V. Instruction bits from the SPI are routed into these buffers. All bias voltages for the polynomial nonlinear capacitors distributed along the nonlinear waveguide are set to 0.9 V. In the experiment studying the expansion of features in data tokens encoded as pulse trains with varying parameters, waveforms were generated using a Tektronix AFG3102C arbitrary waveform generator (1 GSa/s maximum sample rate). The full menu of 125 pulse-train token is derived from combinations of A. βTok amplitudes—0.4, 0.55, 0.65, 0.75, and 0.85 V and (optionally) 1 V—to enable regimes where the coupling is off, partially active and fully active, B. five pulse-train durations TTok(40, 60, 80, 100, and 120 ns) for easy observation of spectrograms and C. five pulse repetition frequencies FTok (20, 40, 60, 80, and 100 MHz). Our preliminary experiments have found that parametric coupling pulses of frequencies higher 150 MHz do not give observable advantage in terms of feature extraction (Supplementary Section 1 of main ref. 21). In this paper, we synthesize analog pulse-train tokens in .csv format and read them off the DAC in the Arbitrary Waveform Generator that injects them into the parameter/token port of the MNN through SMA cables and on-chip Ground-Signal-Ground waveguides. Each measurement consisted of a pair of tokens preceded by a marker and 400 ns dead times before and after injection. After downconversion and optional gain stages off-chip, the analog waveforms were captured using a Tektronix TDS oscilloscope (80 GSa/s maximum). For this experiment, the sampling rate was set to 8 GSa/s to observe a time window of up to 1 μs while maintaining sufficient resolution ( > 125, 000 points per trace). Data acquisition was automated in MATLAB, and spectrograms were computed in Python, with all pair-wise token responses cached for offline Association Matrix construction. After mapping waveform spectrograms to preferred submarine states in the turn-based navigation game, the experiment was repeated in a closed loop. In each rally, two tokens were injected; the MNN’s preferred next state was inferred using the trained linear layer (stored as a .pth file), and the predicted submarine state was applied before the next pair of responses arrived. The cycle then continued, with each new MNN response fed back into the model to determine the subsequent move. The effects of expanding each sub-100-MHz token’s parameters across several gigahertz are clearly visible in the spectrograms, enabling reliable extraction of the appropriate time-frequency slice. For diagnostic purposes, the output microwave spectrum is measured using a Rohde & Schwarz FSU26 spectrum analyzer connected to the output port of waveguide C, which carries a copy of the MNN output signal. This measurement is used to observe the MNN’s instantaneous comb structure and to select the appropriate local-oscillator (LO) frequency from the Anritsu 69347B source for downconversion. Choosing the correct LO ensures that the several-gigahertz-wide bandwidth appearing between 10 and 13 GHz—containing the redistributed spectral features produced by the MNN-is translated into the intermediate-frequency range from DC to 3 GHz. A similar setup is used for the experiment that drives fast bitstreams at 2.5 Gb/s directly into the waveguide ports to produce probabilistic bits. There is no input applied to the token port. Instead, an Anritsu MP1763C pulse pattern generator (PPG) provides 8-bit patterns. These are fed to a waveguide port directly. The oscilloscope is configured to acquire at 625 MSa/s (one-quarter of the bit rate) and is externally triggered on the clock edge from the PPG. In post-processing, only the two down-converted baseband samples that fall within the 3.2 ns window corresponding to each 8-bit sequence are retained. Thresholding and the resulting Static/Dynamic p-bit distributions are computed offline. Each 8-bit binary pattern is injected into the MNN 300 times, and only the baseband output is recorded. This procedure is repeated for every LO frequency in 400 MHz steps, and then repeated again for the next 8-bit pattern, until all 256 patterns are characterized. All collected baseband waveforms are provided in the accompanying datasets. For the LO = 10.4 GHz condition used for the results in the main article, the two-sample baseband waveforms-overlaid across 300 repetitions for each of the 256 eight-bit patterns-are shown as analog eye-diagram-style plots in Supplementary Figs. 13 to 24. The diversity in Static Ratio for all 8-bit patterns is summarized in Supplementary Fig. 4. Importantly, the pulse-pattern generator that produces the high-speed bitstreams outputs each 8-bit sequence cyclically. Although acquisition is triggered by a clock, the capture window may begin at an arbitrary point within this repeating 3.2 ns sequence. Because the bitstream wraps around continuously, the recorded response still corresponds to the action of the same 8-bit symbol striking the MNN, even if the sampling window begins mid-cycle. Consequently, using an estimator such as the Static Ratio, which is sensitive to transition structure and sparsity rather than the exact ordering of bits, is sufficient for capturing the correct statistical behavior of the symbol. For the digital dithering baseline shown in Fig 5c and Supplementary Fig. 12, we numerically generate ideal 8-bit waveforms and analyze them using the same Static/Dynamic criterion. For each of the 256 possible 8-bit binary patterns, we construct a piecewise-constant waveform with high and low levels (1 V/0 V) and 80 samples per bit (640 samples total). We then emulate all possible timing alignments between the sampling clock and the underlying bit pattern by circularly shifting this waveform by every possible sample offset (0–639). For each shift, we read out two samples (a middle and end sample), apply a 0.5 V threshold to convert them to 0/1, and classify the outcome as Dynamic (10 or 01) or Static (00 or 11), exactly as in the experiment. The Static Ratio for each bit pattern is defined as the fraction of shifts that produce a Static outcome. This procedure yields a purely digital prediction of Static Ratios as a function of bit pattern, which we use as a comparison to the MNN-based measurements (with all labels, Static Ratios, and sorted look-up tables saved for analysis). Modeling feature extraction mechanisms through the MNN’s nonlinear dynamics Supplementary Fig. 8 shows the circuit model common to both modes of feature extraction. Linear waveguides (B, C, and D), which are tunable transmission-line stubs31. In this study, however, their lengths are kept fixed and are simply modeled as tank circuits. Nonlinear waveguide A is represented as a cascade of fixed-length inductors, interspersed by nonlinear capacitors. The capacitors, actually constructed from an arrangement of antiparallel diodes32 (sub-panel a.ii), have values that depend on the incoming signal. In effect, these C-L-Cπ nonlinear resonators present time-varying impedances and impart rich nonlinear dynamics due to multiple interferences and reflections along the length of the waveguide. Naturally, these are lossy and must be compensated for. Here a pair of cross-coupled power N-type metal oxide semiconductor (NMOS) transistors provide regenerative, saturable gain given by where G0 is the constant small-signal gain, v is the amplitude of the signal at the drain of the transistor and vsat represents the amplitude at which the gain saturates. Since the linear resonators in the left half (B and C) of the circuit are far-detuned ( ≈ 50 GHz) from the nonlinear resonances’ operating frequencies of around 10 GHz), we can neglect their role in capturing the essential dynamics, and focus on the interactions occurring between waveguides A and D. The two coupling paths are through the parametrically-modulated switch modeled as time-varying capacitor, and a fixed delay path through the coupler. The tunable capacitance transitions between effectively open and short-circuit states, with coupling strength controlled by token amplitude (βTok), pulse frequency (FTok) and pulse-train duration (TTok). We have found that a coupled-mode formulation33,34, shown graphically in sub-panel c.i, is sufficient to capture the underlying processes of the MNN’s instantaneous bandwidth modulation and redistribution of broadband inputs’ features across its output’s spectrum. The nonlinear modes have amplitudes vj (j = 1, 2, …, 6). The linear resonator’s mode amplitude is vlin. The nonlinear modes are coupled through cNL and with the linear mode through the time-varying pulse-token’s coupling strength cpar(βTok, FTok, TTok) and through the passive phase-shift through the Lange coupler cpasseiπ2. Internal losses, due to low on-chip Quality-Factors are γNL,jint and γlinint. The polynomially varying nonlinear resonators’ capacitances take the form CNL,j≈CNL,01−κ1vj+κ2vj2−κ3vj3…. Therefore, their corresponding resonators’ instantaneous resonant frequencies are Taylor-expanded as Adding the terms corresponding to the fast external bit-string B1 and saturable gain, G, the dynamics of the first nonlinear resonator can be compactly written as Mode amplitudes further down the nonlinear waveguide have amplitudes that vary as The linear waveguide mode (D) can be described by These equations are solved numerically with SciPy-based Runge-Kutta-Fehlberg solver35 in Python. The comb bandwidth modulation observed through steady-state Fast Fourier Transforms (Supplementary Fig. 8d) matches observations from the experiment (main Fig. 2b). To simulate how the phase offset of the fast-bit drive influences the MNN’s RF oscillation and produces probabilistic p-bits at one-eighth the bit-speed (main Fig. 5c), we disable parametric modulation and apply only the high-speed token drive. The fast-bit frequency is set to approximately a tenth of the MNN’s RF nominal oscillation frequency. Fixed random initial conditions are used, and a single pass of the 8-bit pattern is injected into the coupled-mode system. The output v1 is mixed with a sine wave of the same frequency as the strongest component of the free-running MNN to produce a baseband signal. It is followed by zero-phase FIR low-pass filtering (801 taps, 312 MHz cutoff, and a fixed +20 dB gain. The filtered waveform is then sampled at FADC = 1/(4Tbit), yielding two analog samples and, effectively, 1 p-bit per pass of the 8-bit input. Generation of the token Association Matrix relating microwave pulse-train tokens Elements of the Association Matrix C quantify how much information about a first token a can be identified from the spectrogram of the MNN’s response to a second token b when both are applied in sequence. The dataset contains one measured example for each ordered pair (a → b) with a ≠ b, giving 15, 500 total examples. Each example contains the class labels a, b ∈ {0, …, 124} and the MNN spectrogram, from which only the time-frequency slice corresponding to the second pulse is extracted from the time-domain response, and flattened into a feature vector Vb∈Rd. The dataset is randomly split into two halves: a training set (XTr,YTr) and a held-out test set (XTe, YTe). Each test example corresponds to a unique ordered pair (a → b), where a is the label of the first token and b is the label of the second token in the original input sequence. There is exactly one test example for each ordered pair (a, b). After training, each fold-specific model f(k) is evaluated on the corresponding held-out test subset (XTe, YTe). Here, a^(k)∈{0,…,124} denotes the discrete class label predicted by the k-th model for the first token, and δa,b(k) simply records whether this predicted label exactly matches the true label a. For the test sample corresponding to the ordered pair (a, b), correctness of the k-th model is recorded as The association score is then defined as the average correctness over folds: This yields a graded measure, in 10% increments, of the fraction of cross-validation folds in which token a remains identifiable from the spectral response associated with token b for the ordered input (a → b), as shown by the colormap in Fig. 3c.ii. Furthermore, the mapping between pulse tokens is generally non-reciprocal, i.e., C[a, b] ≠ C[b, a]. Thus, when the association matrix is used to infer state transitions in a turn-based game or during sentence construction, the order of the tokens is crucial. State compatibility matrix and mechanics of the submarine navigation game The two-player submarine game is conceptually based on the racket sport of squash. Each submarine has a finite length and a limited travel speed, and attempts to intercept a propagating beam of sonar signal that reflects off the water surface, tank walls, and the opposing submarine. In normalized units, the arena is a 100 × 100 grid with two submarines of length L = 20. The sonar beam is treated as a bouncing projectile with an effective radius r = 2 units and travels with a normalized speed v = 1 sea-depth/rally. Its reflections off surfaces are perfectly elastic. In the first rally of each scenario, the beam is launched with a random initial direction between − 45° and 45°. Whenever the beam bounces off either submarine, its horizontal direction reverses. The new horizontal component of the reflection angle depends on where along the submarine’s length the beam made contact. That is, if it hits near the left or right edge, it bounces away with a steeper slope. For simplicity, we compute the reflection angle as Here, x(t) is horizontal position at impact, with xcen at the submarine’s center. The factor π/3 limits the maximum reflection angle. If the above expression exceeds this limit, the angle is simply saturated at ± 60°. The resulting horizontal velocity is then vx=vcos(θ). We also apply a small random horizontal offset to the submarine’s location to prevent perfectly deterministic gameplay. If no collision occurred and the beam reaches the sea-bed, the game ends. We record ordered successful state transitions of the form (ϕA(n)→ϕB(n)→ϕA(n+1)) where ϕA(n) and ϕB(n) denote the discrete states of submarines A and B during successive rally-to-rally transitions. From 10, 000 independent game simulations, we accumulate a compatibility matrix S∈R75×75, with S[b, a] equal to the number of successful rallies in which submarine A assumed state a and submarine B then assumed state b. The matrix is column-wise max-normalized to form a probability estimate This matrix captures behavioral tendencies of a rational defender: given a state choice by submarine A, the entries S~[⋅,a] reflect which responses by submarine B tend to continue the rally. As Fig 3c.i shows, in general, S~[b,a]≠S~[a,b], implying asymmetries in signal interception strategies. Training and closed-loop inference for the turn-based submarine navigation game Discrete submarine states (x, y, θ) span the arena, which is subdivided into 25 spatial bins and 3 angle bins. These 75 states, indexed {0,…,74}, are mapped to (βTok, FTok, TTok) triplets—corresponding to amplitude, pulse frequency, and pulse-train duration, using a greedy search. Submarine B’s behavior is now governed by the MNN. During a rally n, whenever the beam bounces off the MNN’s submarine, we obtain the associated (βTok, FTok, TTok)MNN,n triplet for that event. Submarine A serves as a perfect player with full information about the environment, and when it intercepts the MNN’s transmission, its current state is encoded as (βTok, FTok, TTok)A,n. This ordered pair ((βTok, FTok, TTok)MNN,n, (βTok, FTok, TTok)A,n) is fed to the MNN, and its response is recorded. The corresponding spectrogram slice VMNN,A(n) ∈R is fed to a trained linear layer g:Rd→{0,…,74} that predicts ϕ^MNN(n+1)=g(CMNN,A(n)), representing the next state for MNN-guided submarine. It specifies a desired submarine location (y^MNN,x^MNN) that approximates the optimal next position to intercept the sonar beam. The MNN-controlled submarine then tries to move directly to this predicted position before the incoming signal arrives. Gameplay continues until failure, defined as either (i) the beam missing the MNN-guided submarine or (ii) a commanded movement that exceeds the allowed speed constraint. The total number of successful interception events provides a quantitative measure of survival time for the MNN-guided submarine. The training pipeline begins by mapping physical MNN pulse-train parameters (amplitude, frequency, and duration) to discrete token labels representing submarine states defined by quantized (X, Y, angle of reception) bins. Using this mapping, pairs of experimentally measured MNN spectrograms-each corresponding to two sequential tokens-are extracted and concatenated to form input feature vectors (21 × 20 per token, flattened to 840 dimensions). These spectrogram inputs are used to train a supervised model, a linear-layer neural network, that compares how spectrally similar each MNN spectrogram state is to another and the possibility of one state following the other in sequence (only a two-token context window). Training is performed using cross-entropy loss with Adam optimization and 10-fold cross-validation to select the best-performing model, which is then stored as a .pth file. During inference, the trained linear model is deployed in a closed-loop experiment or simulation. For each rally, tokens corresponding to states of the opposing submarines are injected into the MNN to retrieve the corresponding spectrograms, which are concatenated and flattened into the same 840-dimensional feature vector used during training. This input is passed through the linear layer to produce scores over the 75 possible states, and next, the predicted class is obtained via argmax. The predicted class is then decoded back into (X, Y, angle) bins to determine the next system state, and this process is repeated for each rally (assuming the MNN-guided submarine/drone assumes a viable state that continues the game), enabling evaluation through survival time in the turn-based game. To evaluate robustness, gameplay is repeated over many independent rallies and under multiple limits on submarine movement speed. Histograms of the number of successful MNN-guided interceptions before failure characterize the closed-loop performance of the MNN as a real-time controller, using the saved Python model (.pth file) for inference on spectrogram features derived from oscilloscope waveforms at each rally (see Fig. 4b). Backend model training was done on a single machine with an AMD Ryzen 9 9900X CPU with 64GB of RAM, housing a GeForce RTX 5090 graphics card. In the absence of the physical chip, readers can simulate the full pipeline using pre-recorded responses provided in the accompanying dataset and the modified code. Image quality metrics Here, we formulate the metrics used to quantify perceptual image quality for reconstructions generated from MNN-dithered p-bit outputs, as shown in Supplementary Figs. 11 and 12. These can be plotted using the accompanying code and dataset. Let I(i, j) denote the reference image intensity at pixel (i, j) and Ir(i, j) denote the corresponding reconstructed image. Images are normalized to the range [0, 1] prior to evaluation. N is the total number of pixels. Pixel-wise accuracy is quantified via the mean squared error (MSE), and the corresponding peak signal-to-noise ratio (PSNR) is calculated as where L is the maximum possible intensity value (L = 1 for normalized grayscale images). To capture perceptual and structural fidelity, we attempted to use the structural similarity index (SSIM36). For a local window, let μI and μIr denote the mean intensities of I and Ir, σI2 and σIr2 their variances, and σIIr their covariance. SSIM is defined as where C1 and C2 are small positive constants ( ≈ 0.0005) that stabilize this estimation. The global SSIM score is obtained by averaging over all local windows. SSIM assumes spatial smoothness and structural coherence, but dithering introduces noise to eliminate uniform bands. Because these goals are opposite to one another, SSIM becomes unreliable under dithering, as demonstrated in Supplementary Fig. 11c.ii. We also quantified the similarity of the Fourier magnitude spectra37 of the reference and reconstructed images. With F{⋅} as the two-dimensional Fourier transform, and corr( ⋅ , ⋅ ) as the Pearson correlation computed on vectorized arrays, we first compute X=F{I} and Y=F{Ir} and apply logarithmic compression to reduce dynamic range, The spectral correlation is then defined as where X^ and Y^ are vectorized prior to computing Pearson correlation. This metric verifies that features from the broadband input are redistributed across the entire output spectrum by the MNN in a way that remains recoverable through the dithering scheme. Lastly, to assess the similarity of edges and contours, we computed the gradient magnitude38 of each image. Let ∇x and ∇y denote discrete derivatives along the horizontal and vertical directions, respectively. Using the reference image I and reconstructed image Ir, the corresponding gradient-magnitude maps are The gradient-magnitude correlation is then defined as where the Pearson correlation is computed on the vectorized gradient-magnitude maps. Supplementary information Supplementary Information Transparent Peer Review file