Microwave diffractive neural network chips for sensing and computing Authors: Lei Chen, Qian Wen Wu, Ze Gu, Qian Ma, Lei Xiao, Zhao Tian, Wen Qi Su, Hao Yang Cui, Jian Wei You, Tie Jun Cui Published: 2026 ABSTRACT Electromagnetic diffractive neural networks (DNNs) enable ultra–low-power, low-latency artificial intelligence (AI) inference, yet optical implementations suffer from fabrication and scalability limits, and metasurface microwave systems remain bulky. We present a chip-scale microwave diffractive neural network (MDNN) fabricated in a GaAs semiconductor process, integrating cascaded couplers and phase shifters to implement a diffraction network within a millimeter-scale footprint. The MDNN chip reduces the size of conventional MDNNs by over four orders of magnitude, achieves a computational latency of 2.05 ns, and delivers a system-level energy efficiency of 0.83 TOPS/W. We demonstrate its versatility through three functional prototypes: MNIST handwritten digit recognition, multi-user interference suppression, and real-time obstacle perception for drones. These experiments achieved more than 86% accuracy, validating the capability of the MDNN chip to directly perform both digital image processing and in-situ electromagnetic information processing in the microwave domain. We hope this chip architecture opens a new pathway toward highly integrated designs for electromagnetic DNNs. INTRODUCTION The rapid growth of artificial intelligence (AI) models has imposed unprecedented demands on conventional electronic computing, pushing the traditional architectures toward the limits of fundamental physics and energy efficiency. Photonic integrated circuits, leveraging light as an information carrier, have emerged as a promising platform for AI acceleration, offering inherently parallel processing, high bandwidth, and ultralow power consumption. Optical neural networks (ONNs), including planar and three-dimensional diffractive architectures, have advanced from basic logic operations to large-scale and highly parallel computations (1–7). In particular, diffractive deep neural networks (D2NNs) exploit multi-layer diffraction to perform feed-forward transformations directly on the optical fields, enabling high-throughput tasks such as image classification, phase retrieval, and multi-dimensional data processing. The planar ONN implementations, including fixed nanostructure-based networks, Mach-Zehnder interferometer (MZI) arrays (8, 9), and micro-ring resonators (MRRs) (10, 11), offered reduced device footprint, programmability, and wavelength-multiplexed parallelism, facilitating online and in-situ optical information processing (12–14). Despite these advances, applying ONNs to microwave-domain signals remains challenging. Direct electro-optical conversion introduces significant latency and energy overhead, which undermines the speed and efficiency advantages of the optical computation. In contrast, microwave neural networks (MNNs) retain the benefits of light-speed computation and high parallelism, and enable real-time processing of electromagnetic (EM) wave signals (15, 16), making them suitable for applications such as image recognition and wireless communication encoding and decoding (17, 18). Early MNNs based on passive transmissive metasurfaces (19, 20) rely on fixed phase modulation and lack reconfigurability (16, 21, 22). To overcome this limitation, recent studies have introduced tunable components such as varactors (17, 23, 24) and amplifiers (25–27) to enable dynamically programming of network weights, allowing for online training and multi-task switching (18). Furthermore, by incorporating feedback loops to control amplifier gain, nonlinear operations can be effectively introduced into MNNs. These developments demonstrate the potential of MNNs (17) in terms of performance and flexibility, paving the way for their deployment in EM control applications. Nevertheless, practical deployment of MNNs is limited by their physical size, which restricts its integration into mobile and embedded platforms. Although operating at higher frequencies [such as terahertz (28)] can reduce the wavelength and thus reduce the size, high-frequency devices suffer from large signal loss. More importantly, they are incompatible with some applications in the microwave band, such as wireless communications and automotive radar. Therefore, simply increasing the frequency to reduce the size is not an ideal solution. Consequently, there is a pressing need for highly miniaturized chip-scale MNNs that preserve the computational capabilities and reduce the footprint and power consumption. To solve the problem, here we introduce a microwave diffractive neural network (MDNN) chip implemented using the GaAs semiconductor technology. The chip is integrated with cascaded couplers and phase-shifter arrays within a compact network. By replacing the conventional transmission line with lumped-element couplers and phase shifters, the proposed design achieves a highly miniaturized structure. Through a series of experimental validations, including the MNIST handwritten digit classification, multi-user interference suppression, and UAV-based distance perception, we demonstrate that the MDNN chip supports real-time and in-situ sensing and signal processing with ultralow latency and power consumption. These results provide a scalable foundation toward highly integrated DNN systems, enabling their applications in mobile intelligent system. RESULTS MDNN chip design A pivotal challenge in DNN applications lies in achieving substantial miniaturization of the network size. In the optical domain, elevating the operational frequency has been validated as a direct approach to scale down the metasurface-based diffractive structures, thereby reducing the overall network size. This has been successfully shown in both all-optical diffractive networks and photonic integrated circuits. However, this size-reduction strategy is not suitable for microwave signals due to their inherently longer wavelengths. To overcome this limitation, we apply the concept of lumped-element circuits to MDNN. By translating the wave-domain diffraction structure into an equivalent integrated circuit composed of lumped components, the physical dimensions of the network can be dramatically reduced while preserving its EM functionality. Based on this principle, we realize the MDNN using integrated-circuit technology by applying the lumped-element analysis method. The chip consists of a coupler array followed by phase-shifters (see notes S1 and S2 for detailed description), which are constructed by passive metal interconnects, spiral inductors, and capacitors. This structure provides a direct physical implementation of DNN within a compact chip-scale footprint, as shown in Fig. 1, in which the diffraction structure and hundreds of EM neurons (i.e., the phase modulation units) are integrated within a ten-millimeter size. Compared with the traditional microwave diffraction structures (1, 18), this design significantly reduces the size, computational latency, and power consumption, while achieving a similar neuron density (8) compared to the optical DNN chips. As illustrated in Fig. 1, A and B, we present the designed coupler matrix and phase-shifter matrix. A 3-dB coupler is realized using a π-type network with optimized equivalent inductance and capacitance, and multiple units are cascaded to form the coupler matrix. Numerically optimized performance of the coupler is summarized in Fig. 1, C and D. At the center frequency of 5.8 GHz, the reflection coefficient (S11) at all four ports remains below −20 dB, with an insertion loss of approximately −4 dB. Architecture and core component implementation of the MDNN chip. (A) Lumped-element 3-dB coupler based on a π-type LC network. (B) 3-bit phase-shifter unit, providing eight distinct phase states. (C and D) Simulated S-parameters of the coupler, showing <−20 dB reflection coefficient and around −4 dB transmission coefficient at the central frequency of 5.8 GHz. (E and F) Simulated performance of the eight phase-shifter states, demonstrating a phase step resolution of approximately 45° and insertion losses better than −4.3 dB at 5.8 GHz. (G to I) Conceptual illustrations of three representative applications: handwritten digit recognition, interference suppression in multi-user communication, and UAV-based distance perception. Similarly, by finely tuning the parameters of the π-network, we construct a 3-bit phase-shifter unit capable of providing eight distinct phase states, as shown in Fig. 1E. The detailed design methodology is elaborated in note S2. The corresponding performance of the eight phase-shifter states is shown in Fig. 1, E and F. At the center frequency, the phase difference between adjacent states is about 45°, while the transmission coefficients exceed −4.3 dB, indicating low insertion loss across all states. The measured performance data for both components are provided in notes S2 and S4. This configuration can be mathematically modeled (note S5) as a cascade of coupler and phase-shifter matrices in the microwave regime, which is previously validated in planar network architecture (15). The coupler matrix and phase-shifter matrix can be cascaded to increase network depth and computational capacity. To verify the computational capabilities of the MDNN chip, we designed and trained three DNN chip networks for three typical application scenarios. Firstly, an MNIST handwritten digit classification task was constructed (Fig. 1G), achieving a recognition accuracy of 90%. More significantly, akin to board-level DNN implementations, the proposed chip can directly process the EM domain signals. Accordingly, in the second experiment, we developed a multi-user communication system (Fig. 1H), enabling ultra-low-latency user identification and interference suppression. Thirdly, to further demonstrate the advantages of the MDNN chip in miniaturization, lightweight integration, and power efficiency, we deployed it for real-time obstacle perception on a UAV platform, where the MDNN chip processes the microwave echo signals and generates distance dependent responses with low inference latency (Fig. 1I). These results confirm that the MDNN-chip can perform in-situ EM signal processing in a compact hardware platform. Experiments To evaluate the computational performance of the MDNN chip, we firstly conduct an image recognition task based on the MNIST handwritten digit dataset, as illustrated in Fig. 2. Here, Fig. 2A shows the schematic of image processing system architecture using the MDNN chip. To match the 30 input ports of the MDNN chip, each MNIST image was encoded into a 1 × 30 complex-valued vector by selecting representative low-frequency components from its Fourier domain representation, which was then used as the multi-channel RF input of the MDNN chip. The detailed k-space indexing rule, signal normalization, resolution treatment, and possible physical data injection schemes are provided in note S6. After diffraction processing through MDNN, the output is collected at specific ports (Ports 5, 11, 19, and 26), detected by instruments or detectors, and finally classified. The chip is fabricated using GaAs semiconductor technology, as shown in Fig. 2B. It consists of a coupler matrix and phase-shifter units integrated on a PCB substrate via surface mounting and wire bonding, forming a fully connected 5-layer network (including three hidden layers, each with 30 neurons). The input and output layers contain 30 and 4 nodes, respectively. The chip was tested and validated through patch connectors. A front view of the chip is shown in Fig. 2C, with an assembled size of 26.5 × 16.8 mm2. MDNN was trained using a hardware-aware and quantization-aware workflow, where the measured propagation matrices were used as fixed operators, and the trainable weights were constrained to eight experimentally characterized phase shifter states. The optimized discrete states were then directly mapped to the corresponding passive phase shifter dies during assembly (notes S7 and S8). Based on the actual measured results from the chip shown in Fig. 2C, the confusion matrix for the MNIST validation set is presented in Fig. 2D, achieving an accuracy of 90.03% (see Materials and Methods for more details). Because of the wire bonding area limit and the accumulated loss in discrete die cascading, the fabricated prototype was experimentally demonstrated on four-digit categories. The 10-class MNIST result is therefore used as a numerical scalability validation, since its hardware realization requires a deeper 7-layer network that is currently limited by insertion loss, bonding-wire parasitics, and fan-out constraints (notes S7, S9, and S14). To validate the scalability of the architecture, we numerically simulated a larger-scale integrated network, which achieved 91.47% accuracy on full 10-class MNIST dataset. The confusion matrix for this validation is shown in Fig. 2E, and the corresponding MDNN-chip configuration used in the simulation is provided in note S7. Computational performance of the MDNN chip on MNIST handwritten digit classification. (A) System schematic for image processing. Original images are encoded into 1 × 30 complex-valued vectors through Fourier domain feature extraction, processed by the MDNN chip, and classified via specific output ports. (B) Photograph of the MDNN-chip measurement on a PCB substrate via surface mounting and wire bonding. (C) Front-view photograph of the assembled chip using GaAs technology, with a total size of 26.5 × 16.8 mm2. (D) Confusion matrix showing the classification performance on a four-category MNIST validation set based on the actual chip measurements. (E) Confusion matrix from numerical simulations of a larger-scale network for 10-class MNIST classification. (F) Comparison of computational throughput and system-level energy efficiency between the MDNN platform and conventional CPU/GPU platforms. Compared to previous metasurface-based DNNs (1, 18, 22), the proposed design exhibits a more compact form factor, lower power consumption, and higher computational speed due to the shorter diffraction path. In Fig. 2F, we compare the computational throughput and energy efficiency of the MDNN chip with those of GPU and CPU platforms. The reported 0.83 TOPS/W is evaluated using a unified system-level accounting method that includes the RF source, channel-wise amplification/modulation branches, and output power detectors. The evaluation method is described in note S12. Even at the current limited scale, the forward computation capability and energy efficiency of our chip are already comparable to those of CPU and GPU. With further scaling (see Materials and Methods for cascading details), the energy efficiency is expected to improve significantly. Beyond the computational performance demonstrated above, the most distinctive advantage of the MDNN chip lies in its ability to perform in-situ processing of analog signals directly in the EM domain. To validate this capability, we design a wireless communication scenario with multi-user identification and interference suppression, as illustrated in Fig. 3A. In this setup, four independent users transmit signals through four separate antennas, while the receiver captures these transmissions via four corresponding receiving antennas. In a typical environment, each receiving antenna may pick up the signals intended for other antennas, resulting in inter-user interference. The conventional communication systems often rely on digital AI processing in the backend to mitigate the interference, which inevitably increase the system power consumption and computational overhead. In contrast, the proposed MDNN chip can perform the computational task at the front-end of the demodulation system in real time, achieving low power consumption (see Materials and Methods for more details). Multi-user interference suppression using the MDNN chip. (A) Experimental setup for multi-user signal separation (left). Four independent users transmit signals via separate antennas; the received signals are processed in real time by the MDNN chip for interference cancellation. A photograph of the chip is shown on the right. (B and C) Without the MDNN chip, the measured intensities at the four output ports (B) and the corresponding cross-talk matrix (C) indicate strong inter-port mixing. (D and E) With the MDNN chip, the measured intensities (D) show enhanced port selectivity, and the corresponding cross-talk matrix (E) exhibits power concentrated along the diagonal elements with suppressed off-diagonal interference. (F) Constellation diagrams and corresponding error vector magnitude (EVM) for the four ports under different interference conditions (no channel mixing, single-channel mixing and three-channel mixing), measured with and without the MDNN chip. As shown in Fig. 3A, the received signals are firstly routed to the MDNN chip, where the integrated diffraction network performs real-time interference cancellation. A photograph of the actual chip is presented on the right side. In the absence of MDNN chip (Fig. 3, B and C), the received energy is scattered across all ports without selectivity. The desired signal channels (solid lines), which mean the intensity from relative input to output, are indistinguishable from the undesired interference channels (faint lines). The cross-talk matrix shown in Fig. 3C exhibits a uniform distribution of intensity, indicative of strong signal mixing. In contrast, when integrating the MDNN chip (Fig. 3, D and E), the four target channels achieve higher transmission amplitudes and the cross-talk matrix (Fig. 3E) reveals the energy concentration along the diagonal elements, while the inter-channel interference are suppressed. To further validate this improvement, we perform a communication experiment and the test constellation diagrams are shown in Fig. 3F. With the MDNN chip, the received constellations remain clean and well clustered under all three transmission conditions: no channel mixing, single-channel mixing, and three-channel mixing. We use 32QAM modulation in the wireless communication transmission, where the corresponding error-vector-magnitude (EVM) values of ports 1–3 remain approximately 0.08–0.097. In contrast, removing the MDNN chip yields relatively clean constellations in the no-mixing case, while in three channel mixing a sudden increase of EVM to 0.3 is observed due to the complete interference. For port 4 (CH4), since its relatively-low transmission intensity compared to the other three channels (listed in Fig. 3D), we perform the communication within QPSK modulation and achieve the EVM of about 0.2 in the single or three channel mixing. The increase of EVM values is attributed to the extra path loss of CH4 in MDNN chip. A Monte Carlo sensitivity analysis further indicates that the phase state variations dominate the EVM degradation, and the experimentally observed CH4 EVM of approximately 0.2 arises from the combined effect of baseline network error, fabrication-induced phase/amplitude variations, and residual port reflection or transmission matrix mismatch (see note S13). Despite this, the system quickly deteriorates and exhibits severe distortion when the multi-channel interference is introduced. Under three-channel mixing, the constellation points collapse into noise, and the EVM increases by more than an order of magnitude for all ports. This comparison highlights that the MDNN chip can suppress the interference of the system and enable robust multi-user separation in real time. The most significant advantages of the MDNN chip are its miniaturization and lightweight design, while possessing information processing capabilities similar to the traditional MDNNs and low power consumption, which makes it naturally suited to the needs of unmanned aerial vehicles (UAVs). To further demonstrate the versatility of the MDNN chip in practical EM sensing tasks, we develop an obstacle distance perception experiment using a small UAV, as illustrated in Fig. 4A. The experiment was focused on the real-time MDNN sensing computation chain on a UAV-related platform, where the obstacle-dependent microwave responses are processed to produce distance discriminative outputs for potential onboard perception. In this scenario, the UAV is equipped with a sensing antenna array that transmits and receives microwave signals during flight, and the reflected signals vary with the distance between the UAV and the surrounding obstacle. The received microwave signals are directly guided into the MDNN chip for real-time discrimination and read out by MCU with the detectors. In contrast to the conventional approaches that rely on backend digital signal processing, our system leverages the MDNN chip to process the EM responses directly and generate a distance perception decision with the minimal post-processing. UAV sensing data acquisition configuration based on the MDNN chip. (A) System architecture: the signals from the sensing antennas are processed directly by the MDNN chip, with the outputs classified by a MCU into different distance states. (B) Schematic of the experimental scenario where UAV equipped with a sensing antenna array navigates near the wall. (C) Design of the sensing antenna array, comprising four transmitting antennas fed by a 1-to-4 power divider and four receiving antennas connected to VNA. (D) Photograph of the static data acquisition setup used for training and validation, with a top view of the sensing antenna. To obtain the microwave echo data from the UAV platform in indoor environment, we construct the electromagnetic data acquisition system, as shown in Fig. 4, B and C. The four output ports of the 1-to-4 power divider are connected to four external transmitting antennas (Tx1-Tx4), while the four internal receiving antennas (Rx1-Rx4) are connected to the vector network analyzer (VNA) to capture the reflected waves from the environment. The reflected signals from the environment vary significantly depending on the distance from UAV to wall. Detailed design of the sensing antenna is elaborated in note S3. In our data acquisition configuration, we conduct a data-acquisition experiment in Fig. 4D, where UAV (with the sensing antenna array) was positioned at representative distances relative to the wall and the reflected waves were recorded by VNA. To accurately collect the S-parameters at the antenna end of the UAV platform, we use a method of suspending UAV with lightweight RF cables to collect the data. It should be noted that the RF cables and VNA were used only for echo acquisition and dataset construction, serving as the laboratory interfaces rather than the fundamental components of the MDNN sensing mechanism. During the distance perception validation, the obstacle-dependent sensing signals were processed by the MDNN chip in real time, and the distance-dependent outputs were obtained after detector readout. Any deviations caused by the UAV’s vibrations and movements are included in the dataset. These measurements, taken in multiple rounds at different distances, form the training dataset for the obstacle distance classifier. Following the training strategy described in note S7, we incorporate the frequency dimension as an additional informative feature for the UAV dataset, enabling the network to use both spatial and frequency signatures in the reflected EM responses without introducing extra backend processing. A global view of the experimental arena is presented in Fig. 5A, with the integrated MDNN chip shown in Fig. 5B. At different UAV-wall distances, the microwave echo responses were processed by the MDNN chip to perform real-time distance state perception, demonstrating its ability to generate the distance discriminative outputs with low inference latency. As UAV approaches the wall from a distant position to a near position, the MDNN chip is able to perform real-time distance perception. The system was evaluated in two indoor environments: a single-wall scenario and a corner configuration. In the single-wall experiment (Fig. 5C), UAV maintains an approximately constant height and moves horizontally along the wall-normal direction. Snapshots taken at representative distances illustrate the corresponding flight states. In the corner experiment (Fig. 5D), we select a vertical corner and translate the UAV across different positions on the floor plane. This setup tests whether the system can still estimate distance reliably when the geometry is more complex and reflections are stronger. MDNN-based UAV distance perception validation and performance evaluation. (A) Overview of the experimental environment, showing the UAV navigating along an indoor corridor. (B) Photograph of the MDNN chip. (C) Snapshots of UAV at representative positions during the distance perception. (D) Experimental scene of UAV operating in a corner environment. (E and F) Simulation and experimental results for the single-wall scenario. When realistic quantization constraints are applied, the model achieves an exceptional AUC of 0.907 and an accuracy of 86.1%. (G and H) Performance in a corner environment. Experimental results yield an AUC of 0.9645 and an accuracy of 86.0% after incorporating the realistic quantization effects. Figure 5, E and F, present the simulation and experimental results for the single-wall scenario. To comprehensively evaluate the classification performance, we adopt the area under the receiver operating characteristic curve (AUC) as a metric. An AUC of 0.907 and an accuracy of 86.1% is achieved after incorporating the realistic quantization effects into the simulation. Similarly, experimental validation in a corner configuration yields an AUC of 0.9645 with an accuracy of 86.0%, as summarized in Fig. 5, G and H. These results demonstrate the MDNN chip’s ability to perform real-time and low-latency EM sensing under indoor conditions, indicating its potential for UAV navigation and proximity awareness with further onboard integration (note S3). DISCUSSION We proposed and experimentally validated a MDNN chip that can perform the digital image recognition and in-situ EM information sensing and processing. Utilizing a GaAs semiconductor process, we firstly designed the integrated circuits for couplers and phase shifters, and introduced a planar-integration scheme. Three chip prototypes are fabricated and tested using the PCB surface-mount and wire-bonding techniques, demonstrating successful MNIST handwritten digit classification, multi-user interference suppression in the wireless channels, and UAV-related obstacle distance perception. Compared to the subwavelength metasurface arrays in the microwave regimes, the proposed MDNN chip achieves a reduction in overall footprint by over four orders of magnitude, a computational latency as low as 2.05 ns, and a system-level energy efficiency of 0.83 TOPS/W for the current 4-class platform (see notes S11 and S12 for details). We believe that the proposed approach not only significantly enhances the performance of MDNNs but also substantially broadens their application scope through miniaturization, laying a foundation for future application-specific integrated devices. Furthermore, for the AI tasks typically handled by digital systems, the MDNN chip offers considerable network scale and parallel computing capacity in a compact form factor. Benefiting from mature RF integrated circuit technology, the MDNN chip shows promising potential to realize large-scale programmable EM AI computing architectures and high-performance nonlinear activation functions, thus offering a competitive alternative to the optical DNN chips (see Materials and Methods for nonlinear capability). The present prototype still has several engineering limitations. It uses passive fixed weights obtained from hardware-aware training, and its wave-domain computation remains linear before terminal detection, making it more suited to the RF feature transformation, interference suppression, channel decoupling, and sensing tasks governed by the EM superposition. Future work will be focused on low-loss layers, monolithic integration, reprogrammable phase shifters, and in-situ calibration/training. MATERIALS AND METHODS Chip cascading and scaling The MDNN chip integration in this work employs the PCB-based mounting and bonding wire cascading approach (Fig. 2B). This method offers high flexibility and cost-effectiveness in the laboratory testing, facilitating the error correction and iterative optimization in experiments. Owing to the batch-processing of chip fabrication, the coupling and phase-shifting layers can be easily expanded via bonding wire, as long as pick-and-place machines allow. However, this approach may introduce parasitic inductance errors from the bonding-wire during fabrication, necessitating iterative calibration during testing. A more advanced and highly integrated approach for the future would involve integrating all chip components onto a single wafer die. This, however, requires ensuring that each structure achieves optimal performance and connectivity within the overall network. Therefore, extensive testing, optimization, and iterations are essential before realizing this goal. Monolithic integration on the same wafer die would significantly enhance system integration, reduce fabrication errors, and simplify chip packaging. For the 10-class MNIST task, the limitation arises mainly from the required 7-layer depth rather than the output port number. With the current discrete die assembly, such scaling would exceed the wire bonding area after RF fan out and introduce excessive cascading loss, bonding parasitic and mismatch, motivating lower loss components and monolithic integration in the future work (note S14). On-chip measurement Each port is designed with a pluggable connector socket functioning as a SPDT switch. A surface-mount resistor is connected to one of the hands. When the external connector is not placed, the corresponding port is substituted by a perfect matching load that absorbs any injected signal. In this way, we could measure the transmission matrix without being influenced by the adjacent ports’ cross-talk or reflection. We use a two-port vector network analyzer (SNA5032A) to measure the S-parameter. Through manually traversing all the port combination, the matrix is collected. Specifically, the microstrip lines of different length between the pads and the connector sockets are de-embedded using the calibration algorithm. The nonlinearity of MDNN chip The current MDNN chip primarily operates as a linear network, and has already demonstrated strong performance in applications such as image recognition, communication interference suppression, and UAV sensing. To isolate the contribution of the diffractive structure, we further compared MDNN with digital linear baseline models and conducted an ablation study, as shown in note S10. There are two feasible routes to enhance its nonlinear capability. The first approach is to append one or more nonlinear digital layers after the linear MDNN, leveraging high-performance edge hardware such as FPGAs. This method introduces minimal additional power or latency overhead and is compatible with existing computational platforms. The second approach is to design dedicated integrated circuit structures that provide intrinsic nonlinear activation. This would enable higher integration density and lower nonlinear processing latency, representing a key direction for future research. The feasibility and requirements of future on-chip training are discussed in note S15. Chip power and computing capacity estimation Currently, the MDNN chip does not include reprogrammable modules, and the passive coupler-phase shifter network itself consumes negligible DC power. The path loss inside the propagation process plays a key role in the power consumption of the MDNN analog platform. A lower propagation decay, together with highly sensitive power detectors, is preferred. Using the commercial power detectors as references, a dynamic range below −50 dBm can be reached without middle-stage amplifiers. Meanwhile, the average insertion loss for each propagation layer and modulation layer can be estimated asL=PoutPin≈tr(THT)N·Lps=15dB(1)where 𝑇∈ℂ30×30 is the measured transmission matrix of one propagation layer, N is the number of spatial channels, and L is the average insertion loss of the phase shifters. For a three-layer MDNN chip, an approximate path loss of 45 dB is therefore anticipated. Under this core level link budget estimate, direct power monitoring without additional middle stage amplifiers requires an RF power of approximately −5 dBm, corresponding to 0.3 mW, at each input port. For 30 input ports, the total RF power delivered to the MDNN input ports is approximately 9 mW. Combining this with four output power detectors operated at 3.3 V and 30 mA, the minimum core-level power budget isP=9mW+396mW=405mW(2) This value reflects the RF input power delivered to the passive MDNN chip and the power required by the output detection modules. When functioning as an analog computing platform, however, the total system power consumption should include not only the RF power delivered to the MDNN input ports and the output detection modules, but also the signal generation and channel-wise amplification/modulation modules. We further evaluate the MDNN-based analog platform using a system-level power budget. The total system power is calculated asP=Psource+Pmod+PPD=0.33W+30∗(40mW+160mW)+4∗100mW=6.73W(3)where Psource denotes the RF source power consumption, Pmod denotes the power consumption of the channel-wise amplification and modulation branches, and PPD denotes the power consumption of the output power detectors. For the current 4-class MNIST MDNN platform, the system includes one RF source, 30 channel-wise amplification/modulation branches, and four output power detectors. Here, 0.33 W corresponds to the RF source, 40 mW corresponds to the amplifier power per channel, 160 mW corresponds to the DAC/modulation branch power per channel, and 100 mW corresponds to the power detector consumption per output channel. With a computing capacity of 5.6 TOPS, the system-level energy efficiency of the current 4-class MDNN platform isη=5.6TOPS6.77W=0.83TOPS/W(4) The detailed component-level power budget, link budget estimation, and comparison with CPU/GPU platforms and other microwave diffractive neural networks are provided in note S12 and table S2. Image classification task To fit the input dimension of the on-chip platform, the MNIST images are first converted into complex-valued RF input vectors by Fourier domain feature extraction. A 2D discrete Fourier transform is applied to each 28 × 28 image, and 30 low-frequency k-space components closest to the center are selected after discarding the DC component, which mainly represents the average image intensity. These selected complex coefficients form the 1 × 30 input vector, where the real and imaginary parts correspond to the in-phase and quadrature components of coherent RF channels. No additional source quantization is applied in the main experiment, and a global input normalization is unnecessary for class decision because the output label is determined by the maximum detected port power. The influence of input quantization and possible physical RF data injection schemes is discussed in note S6. The classical cross-entropy function is used as the loss function to measure the distance between truth labels and normalized output intensity using softmax function. For four-category classification, the sample number of training dataset and validation dataset are 20,000 and 2000 respectively. For 10-category classification, the sample number are 50,000 and 10,000 respectively. In the training, we use the gradient-descend algorithm for complex-value neural networks to accelerate the training. Specific measures are taken to counter the parametric quantization effect and cascading power decay during the forward propagation (see note S7). After training, the optimized discrete phase shifter indices are mapped to the hardware as described above. For experimental validation, the transmission matrix of the assembled substrate circuit is directly measured, and the forward propagation is emulated on a computer to retrieve the output port intensities for classification accuracy calculation and post-processing. Wireless communication task In the goal of compensating the wireless channel using on-chip platform, the loss function in the wireless communication task is designed to minimize the cross-talk while retaining the maximum power at the receiver. We note here that the on-chip platform indeed introduces additional loss in the budget link. However, the overall physical architecture using couplers and phase shifters are ideally unitary, which means further module optimization could approach the non-loss property. Therefore, the proposed wireless communication demonstration still has its merit. The loss function is defined asloss=−sumi∣Λii∣+ksumi≠j∣Λij∣(5)where Λ is the combined transmission matrix between the wireless channel T and compensated matrix Tchip, i.e.Λ=TchipT(6) The training dataset is generated using the commercial electromagnetic software to measure the transmission between the four Tx and Rx antennas. The distance between the adjacent antennas is 10 cm (∼2λ) and the distance between transmitter and receiver is 56 cm (∼2λ). We note here that with the future reprogramming ability introduced in on-chip platforms, the channel compensation could be adaptive against dynamical scenarios. Through deliberately perturbing the Tchip values using noises and geometry error in simulation, we expand the dataset for data augmentation. The training dataset has a size of 10,000, while the validation dataset is the same. We use the gradient-based optimization algorithm to train this regression problem. Similar to the image classification task, we bond the corresponding phase shifters dies with the coupler dies to build the on-chip wireless channel compensation module. The transmission property is again measured to produce the evaluation results. Accuracy rate is defined as the four individual bi-stream transmission for four channels using the OOK modulation with equivalent transmission power. In high-order transmission simulation, 32-QAM and QPSK modulation scheme are adopted for different respectively to calculate the communication performance.