Framework and method for ai-assisted real-time wireless coverage map generator for live wireless network digital twinning

Abstract

A method of estimating a radio map from environmental context comprises designating a training area and for each transmitter location in the training area, generating a channel parameter radio map that is centered on the transmitter and is based on a ray-tracing twin model and generating a two-dimensional elevation map, centered on the transmitter, corresponding to the training area. The method may further comprise training a U-Net-based or any other neural network-based model using the channel parameter radio maps and the two-dimensional building elevation maps corresponding to the transmitters in the area, to produce a trained model, and designating an area to be radio-mapped. The method may further comprise providing the trained with a two-dimensional input elevation map derived from a three-dimensional model of the area to be radio mapped, and producing, by the model, an estimate of the radio map based on the two-dimensional elevation map.

Description

Framework and Method for ALAssisted Real-Time Wireless Coverage Map Generator for Live Wireless Network Digital TwinningRELATED APPLICATION(S)

[0001] This application claims the benefit of U. S. Provisional Application No.63 / 709,827, filed on October 21, 2024. The entire teachings of the above application are incorporated herein by reference.GOVERNMENT SUPPORT

[0002] This invention was made with government support under Grant No. FED-OTH / DOC-NIST 25-60-IF011 / Johari awarded by the National Institute of Standards and Technology. The government has certain rights in the invention.BACKGROUND

[0003] Digital twins have emerged as transformative tools in wireless networks, creating virtual replicas of physical deployments that operate alongside the real world. Real-time digital twins let developers design, test, and evaluate “what-if’ scenarios without risking live operations. Furthermore, by operating ahead of real time, they can forecast potential events, and in a fully closed-loop implementation, continuously ingest live data, update their models, and even trigger actions back in the physical network. Real-time digital twins support a broad spectrum of wireless network use cases by leveraging continuously updated environmental models and fast inference engines.

[0004] While traditional network simulators and digital twins both create virtual representations of networks, they fundamentally differ in their relationship with physical systems. Conventional simulators operate in isolation, running predefined scenarios with static inputs to predict theoretical outcomes. In contrast, digital twin networks establish a bidirectional connection with their physical counterparts, continuously ingesting real-time data to create high-fidelity representations that evolve alongside the actual network. This interactive mapping enables digital twins to not only reflect the current state of the network but also to provide closed-loop automation capabilities, where changes validated in the virtual environment can be safely applied to the physical network. This real-time synchronization capability makes digital twins particularly valuable for mission-critical applications in next-generation wireless networks, where they can facilitate network optimization, predictive maintenance, and innovative service development without riskingoperational disruptions. In modern Radio Access Network (RAN) architectures, the threshold for “real-time” performance varies by control-loop function and application. Non-real-time RAN Intelligent Controller applications operate on timescales exceeding one second for longterm optimization. Near-real-time functions execute within ten to one thousand milliseconds for time-sensitive control. The most stringent tier, distributed applications, enables sub-ten-millisecond response times. Accordingly, digital-twin systems must meet these diverse latency targets, with the strictest real-time requirement - data acquisition and response - set below ten milliseconds for distributed application-level operations.

[0005] To achieve a digital twin framework that is effective across all layers of the protocol stack, accurate and agile channel modeling is essential to precisely and swiftly characterize the radio signal propagation through a dynamic environment. This constitutes the basis of a high-fidelity digital twinning system. Channel modeling serves as the foundational layer upon which all digital twin capabilities are built. Without accurate propagation models, digital twins cannot reliably predict network behavior, optimize resource allocation, or trigger preemptive actions in the physical network. The challenge lies in achieving the dual requirements of high-accuracy and ultra-low-latency, for which channel models must be computed fast enough to support real-time decision making, while maintaining sufficient fidelity to ensure reliable network operations. This creates a fundamental tension between computational complexity and model accuracy that existing approaches struggle to resolve.SUMMARY

[0006] The embodiments described herein are directed to an Al-driven framework for real-time radio-map estimation that bridges the gap between high-fidelity ray tracing and fast, scalable inference. By relying solely on a novel single-channel elevation map - which is simple, cost-effective, and widely accessible - as input, and a U-Net-based autoencoder, the described embodiments generate accurate path gain maps in under 4ms. The data-driven design allows efficient transfer-learning calibration using a small number of field measurements, enabling site-specific adaptation and correction of residual errors. Extensive evaluation on a large Boston-area dataset shows an example embodiment matches measurement-based performance on key system-level metrics, such as spectral efficiency and block error rate (BLER). Integration into the Colosseum emulator and the Sionna SYS platform confirms the practical feasibility of the example embodiment for real-time, end-to-end wireless network emulation.

[0007] A key feature of the described embodiments is the resolution-adaptive input representation: while the model accepts fixed-size elevation map inputs, it supports variable spatial resolutions, enabling radio-map estimation across physical areas ranging from 500m to 3km per side without altering the architecture. This design supports flexible deployment scenarios - from dense urban zones to wider suburban or campus-scale regions - while preserving inference speed and model accuracy.

[0008] The advantages of the described embodiments lie in drastically reducing computational cost and latency for radio environment modeling, while using only easily obtainable elevation maps. Its data driven nature allows for seamless incorporation of realistic wireless channel effects into real-time network management, adaptive resource allocation, and rapid scenario testing. These capabilities empower researchers and network operators to dynamically optimize system performance and accelerate the development of advanced digital twin applications requiring fast, accurate propagation predictions from minimal input data.

[0009] In one aspect, the invention may be a method of estimating or predicting a radio map consisting of a coverage map, or any radio network characteristic, from environmental context, comprising designating a data-driven training area, and for each transmitter location in the training area, (i) generating a channel parameter radio map that is centered on the transmitter and is based on a ray-tracing-twin or any other channel model, and (ii) generating a two-dimensional elevation map, centered on the transmitter, corresponding to the training area. The method may further comprise training a U-Net-based model or any other neural network-based model using the channel parameter radio maps and the two-dimensional building elevation maps corresponding to the transmitters in the area, to produce a trained U-Net-based or any other neural network-based model. The method may further comprise designating an area to be radio-mapped, providing the trained U-Net-based or any other neural network-based model with a two-dimensional input elevation map derived from a three-dimensional model of the area to be radio mapped, and producing, by the U-Net-based or any other neural network-based model, an estimate of the radio map based on the two-dimensional elevation map.

[0010] The channel parameter is path gain. The channel parameter is delay spread. The method may further comprise calibrating the trained U-Net-based or any other neural network-based model using one or more channel characteristic field measurements. The method may further comprise calibrating the trained U-Net-based or any other neuralnetwork-based model using a subset of collected channel characteristic measurements. The method may further comprise modifying a loss function associated with calibrating the U-Net-based or any other neural network-based model using a weighted formulation that emphasizes measured points without overriding knowledge learned from simulation.

[0011] The method may further comprise configuring a set of the two-dimensional input elevation maps to each have a fixed spatial dimension, wherein a spatial resolution from map to map varies within the set. The method may further comprise applying minmax normalization by the U-Net-based or any other neural network-based model. The method may further comprise inverting elevation values within the two-dimensional input elevation map processed by the U-Net-based or any other neural network-based model. The method may further comprise generating the two-dimensional input elevation map by encoding building heights, terrain elevation, and physical obstructions within the area to be radio mapped.

[0012] In one aspect, the invention may be a system for estimating a radio map consisting of a coverage map, or any radio network characteristic from environmental context, comprising a U-Net-based or any other neural network-based model trained using one or more channel parameter radio maps and one or more two-dimensional building elevation maps, a two-dimensional input elevation map derived from a three-dimensional model of an area to be radio mapped. The two-dimensional input elevation map may be applied to an input of the U-Net-based or any other neural network-based model to produce an estimate of the radio map at an output of the U-Net-based or any other neural network-based model.

[0013] The one or more channel parameter radio maps may be generated based on a raytracing twin model, and wherein the two-dimensional elevation map is generated corresponding a training area. The U-Net-based or any other neural network-based model may be trained with the channel parameter radio maps and the two-dimensional building elevation maps corresponding to the transmitters in the area. The channel parameter may be path gain. The channel parameter may be delay spread. The U-Net-based or any other neural network-based model may be trained with a subset of collected channel characteristic measurements. The U-Net-based or any other neural network-based model may be trained with a subset of collected channel characteristic measurements. The U-Net-based or any other neural network-based model may use minmax normalization. The two-dimensional input elevation map may use inverted elevation values.

[0014] In another aspect, the invention may be a method of estimating a radio map from environmental context, comprising designating a training area, generating a channel parameter radio map that is based on a ray-tracing twin model, and generating two-dimensional elevation maps corresponding to the training area. The method may further comprise training a deep learning model using the channel parameter radio maps and the two-dimensional building elevation maps to produce a trained U-Net-based or any other neural network-based model, designating an area to be radio-mapped, providing the trained U-Net-based or any other neural network-based model with a two-dimensional input elevation map derived from a three-dimensional model of the area to be radio mapped, and producing, by the U-Net-based or any other neural network-based model, an estimate of the radio map based on the two-dimensional elevation map.BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

[0016] The foregoing will be apparent from the following more particular description of example embodiments, as illustrated in the accompanying drawings in which like reference characters refer to the same parts throughout the different views. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating embodiments.

[0017] FIG. 1 shows a transmitter / receiver measurement scenario across a geographic region.

[0018] FIGs. 2A and 2B show ray-tracing results - generated using a maximum-detail configuration on the same locations as our measurements -compared to the empirical data.

[0019] FIG. 3 shows material and antenna misconfigurations, Empirical Cumulative Distribution Function in ray tracing.

[0020] FIG. 4 shows the measured path gain values plotted and ordered by distance from the transmitter.

[0021] FIG. 5 shows a geographic study area in Boston, MA.

[0022] FIG. 6 shows the simulation runtime for 100 scenarios across different ray-tracing configurations.

[0023] FIG. 7A illustrates how varying certain parameters affects the resulting path-gain radio maps in a representative scenario.

[0024] FIG. 7B shows a summary of studies that employed convention machine learning techniques for channel parameter prediction.

[0025] FIG. 7C shows a summary of DL-based methods for channel parameter prediction.

[0026] FIG. 8 shows a DL framework for estimating radio maps from environmental context.

[0027] FIG. 9A shows a network structure for estimating radio maps from environmental context.

[0028] FIG. 9B shows architectural details for a network structure for estimating radio maps from environmental context.

[0029] FIG. 10 shows an elevation map, associated U-Net output, and associated high-fidelity ray tracing output.

[0030] FIG. 11 shows the median error percentage across all folds and coverage maps.

[0031] FIG. 12A and 12B show an example embodiment of a calibration pipeline using transfer learning with a weighted loss to align simulation with measurement data.

[0032] FIG. 13 shows the performance comparison with and without calibration, between Sionna RT, the calibrated U-Net model of the described embodiments, and the uncalibrated model.

[0033] FIG. 14 shows what happens when a user selects or updates a location.

[0034] FIGs. 15 A, 15B, and 15C compare errors for three channel -modeling approaches.

[0035] FIG. 16 shows Shannon capacity heatmaps for ray tracing, uncalibrated U-Net, measurement data, and calibrated U-Net model estimations on test measurement points.

[0036] FIG. 17 shows results from running OpenAirInterface (OAI) on the Colosseum testbed.DETAILED DESCRIPTION

[0037] A description of example embodiments follows.

[0038] The embodiments described herein are directed to a deep learning-based framework for real-time, high-fidelity radio map estimation. Channel modeling for RF scenarios traditionally falls into three main categories: (i) measurement-based models, (ii) statistical models, and (iii) deterministic methods (e.g., ray-tracing).

[0039] Measurement-based approaches capture site-specific phenomena with high accuracy but are costly, labor-intensive, and quickly outdated in dynamic environments.

[0040] Statistical channel models employ stochastic or deterministic mathematical equations to characterize wireless propagation, but these models often fail to represent all scenarios or capture environmental intricacies. Their reliance on simplified environmental assumptions leads to prediction inaccuracies, particularly in site-specific scenarios.

[0041] Ray tracing provides a deterministic means of modeling wireless channels by launching rays and simulating their interactions - reflection, diffraction, and transmission -with environmental geometries using material-specific parameters. Although ray tracing achieves higher site-specific accuracy than empirical models — especially in architecturally complex settings — it remains impractical for real-time use. Even with GPU-accelerated implementations, the computational burden of dynamic mobility modeling and scenario adaptation is prohibitive, and the exhaustive precomputation generally demands extensive storage. Furthermore, ray tracing fidelity depends on highly detailed 3D maps with accurate material assignments and accurate antenna patterns, since electromagnetic responses vary noticeably across surfaces. Therefore, routinely, in a trade-off against computational burden, ray tracing remains an approximation — constrained by finite ray sampling and often incomplete modeling of propagation phenomena like diffraction, scattering, and reflection for all possible points. These factors limit ray tracing’s suitability for high-fidelity channel modeling in digital twins.

[0042] A balance between accuracy and computational efficiency may be found through advanced propagation modeling. Deep Learning (DL) excels in this regard by training on extensive propagation datasets to capture complex channel behaviors without explicit geometric simulation. Artificial Intelligence (AI) / Machine Learning (ML) frameworks can automatically learn nonlinear relationships among environmental features, measurements, and spatial dynamics. Consequently, AI / ML frameworks can deliver high-fidelity, real-time channel estimates, generalize across varied scenarios, and continuously refine their estimations as new data arrive - all while sidestepping the computational burdens inherent in conventional approaches.

[0043] Radio (environment) maps provide a two-dimensional representation of averaged statistics of channel characteristics (e.g., received signal power, interference power, power spectral density, delay spread, and channel gain) over a geographic region. Unlike individual point-wise channel estimates, radio maps capture spatial relationships and large-scale propagation patterns, reflecting how neighboring locations influence one another. Thisinherent spatial structure makes radio maps a natural and effective output format for DL models, which can leverage locality and spatial dependencies.

[0044] Convolutional Neural Network (CNN) architectures can process environmental inputs such as terrain and building layouts to generate channel predictions for an entire area in a single inference step, rather than predicting each point independently. This structured approach not only improves scalability and inference speed, but also provides the spatial context essential for real-time digital twin applications. An example embodiment, referred to herein as AIRMap, is directed to a deep-learning-based framework for real-time, high-fidelity radio map estimation. AIRMap is trained on a large site-specific dataset, automatically generated through a scalable pipeline using ray-tracing simulations and 2D elevation data. Unlike prior models requiring multiple inputs, AIRMap leverages a single-channel elevation map to produce accurate channel predictions with sub-5 dB root mean square error (RMSE) and millisecond-level inference latency. A lightweight transfer-learning calibration procedure using sparse field measurements significantly improves accuracy, reducing median error to 10%. AIRMap is integrated into two platforms - Sionna™ SYS (a system-level simulation module) and the Colosseum testbed - where it achieves near-zero error on spectral efficiency and block error rate metrics, validating its suitability for realtime digital twin applications across protocol layers.

[0045] The described embodiments derive from at least the following six innovations.

[0046] Ray-Tracing Efficiency Analysis - The computational complexity-fidelity tradeoffs in ray-tracing simulations are systematically analyzed to identify optimal configurations that enable large-scale dataset generation while maintaining high fidelity for deep learning applications.

[0047] Large-Scale Radio Map Dataset: An automated pipeline generates the largest sitespecific radio map dataset to date, comprising 60,000 Boston-area samples with diverse propagation scenarios for robust neural network training.

[0048] Variable-Scale Coverage Modeling: Each sample covers a square region with side lengths ranging from 500m to 3km, allowing the model to generalize across both local and wide-area propagation conditions.

[0049] Single-Input U-Net Architecture: A novel neural network architecture is employed that requires only 2D elevation maps as input, achieving sub-5 dB root mean squared error (RMSE) with 4ms inference time — over 7000 times faster than GPU-accelerated ray tracing.

[0050] Fine-tuning: A novel calibration framework is used that leverages large-scale simulated ray-tracing data for pretraining, followed by fine-tuning with a small subset of real-world measurements. This approach effectively bridges the simulation-to-reality gap, enabling practical deployment with minimal measurement overhead while preserving model generalization and accuracy.

[0051] Real-Time Testbed Integration: The first sub- 10ms radio map generation in operational wireless testbeds (Colosseum and Sionna™ SYS) is demonstrated, enabling true real-time digital twin applications with near-zero system level performance error.PRACTICAL ANALYSIS OF RAY-TRACING-BASED CHANNEL MODELS

[0052] Presented is an analysis of a measurement scenario to assess the impact of simulation parameters on the gap between ray-tracing predictions and real-world measurements. Since simulation data forms the foundation for training the DL model, a calibration approach is employed that uses a small set of field measurements to fine-tune the model. This enables the data-driven model to surpass ray tracing in accuracy by correcting for systematic simulation errors.Measurement Campaign.

[0053] A measurement scenario was configured to capture channel characteristics at 933 MHz within a section of the Northeastern University campus. In this setup, the transmitter (TX) remains stationary on the rooftop of Dodge Hall, while the receiver (RX) is mobile, following the route depicted in FIG. 1. The measurement campaign map of FIG. 1 shows the Transmitter (TX) location 102 and Receiver (RX) route 104. The line color represents the path gain and coverage distribution as shown by the legend on the right side of FIG. 1.

[0054] For the TX side, we used an Ettus Universal Software Radio Peripheral (USRP™) X410 as the Radio Unit (RU), followed by a Mini-Circuits® ZHL-1000-3 W+ amplifier, and a Pasternack® PE51OM1014 antenna. On the RX side, the same antenna is employed as is used on the TX side. The RX RU was the Viavi Solutions™ T / Rx Software Defined Transceiver.

[0055] For channel measurements, the maximum available bandwidth of 2MHz at a center frequency of 933 MHz is used. Transmission was conducted in bursts of 500 consecutive GLFSR-14 codewords every 0.1 seconds. All losses and gains from cables, amplifiers, antennas, and other hardware components were removed by characterizing the entire setup in an anechoic chamber under identical conditions. A summary of the measurement setup is provided in Table I.TABLE I: Measurement Setup and EquipmentTX RU 1-11 in USRP X410Amplifier Minicircuits ZHL-1000-3W+ (38 dB)Antenna Pasternack PE51OM1014 (6 dBi)Location 42 “20’25”N 71°05’16”WRX RU T / RX provided by VIAVI SolutionsAntenna Pasternack PE51OM1014 (6 dBi)Location Mobile (see Fig. 1 )Meas. DetailsFrequency 933 MHzBandwidth 2 MHzCodeword GLFSR-14Synchronization GPS clock for both TX and RXPath Gain Computation from CIR.

[0056] The path gain was computed directly from the measured Channel Impulse Response (CIR), as shown in equation (1):where ai(t) is the complex amplitude and ri(t) is the delay of the i-th path at time t. The total path gain at each instant can then be computed by integrating the squared magnitude of the CIR, as shown in equation (2):

[0057] In the mobile measurement scenario, the receiver’s location changes over time; thus, each measurement can be associated with the receiver position t / nA / ). Under this setting, the time-indexed path gain P( ) can be equivalently expressed as a location-dependent function P(< / RX), SO that the measured data can be mapped directly into spatially indexed radio maps.

[0058] In FIGs. 2A and 2B, the ray-tracing (Sionna) results - generated using a maximum-detail configuration on the same locations as our measurements - are compared to the empirical data. The 3D environment model was sourced from the Boston Planning Department, with all surfaces assigned concrete material properties and terrain modeled asdry ground. FIG. 2A shows a comparison of measured and ray-traced path gain values across locations. Maximum normalized correlation is 0.3967 at zero lag. FIG. 2A demonstrates that ray tracing and measured path gains exhibit substantially different levels. The scatter plot in FIG. 2B further illustrates that these differences persist across all sample points, deviating from the ideal linear relationship. This discrepancy highlights the limitations of ray tracing’s approximations and its inability to fully capture real-world propagation.

[0059] Calibration techniques that adjust material properties can improve ray-tracing fidelity, but they are typically limited to small-scale indoor settings and do not extend well to large outdoor environments. To quantify the effect of misconfigured materials and antenna patterns on path gain - especially when detailed environmental parameters are unknown -a series of Sionna™ RT simulations were performed. 50 different random scenarios were conducted with a maximum depth of 20 and diffraction enabled. In each run, all objects were initially assigned the International Telecommunications Union (ITU) brick material model, and then the resulting path-gain error was measured assuming the true material was concrete. This procedure was repeated for antenna patterns by comparing TR 38.901 specifications against an isotropic radiator to isolate the impact of antenna misconfiguration. The results, summarized in FIG. 3, illustrate how incorrect material assumptions or antenna-pattern settings degrade path-gain accuracy. FIG. 3 shows material and antenna misconfigurations, Empirical Cumulative Distribution Function (eCDF) in ray tracing.

[0060] In FIG. 4, the measured path gain values are plotted, ordered by distance from the TX. The highly irregular urban geometry prevents a simple radial approximation, as signal propagation is neither symmetric nor uniform. Consequently, site-specific inputs - such as detailed environmental geometry - are essential. Traditional statistical models, which are not tailored to a particular site, fail to capture the complex, nonlinear propagation effects present in urban areas and therefore lack sufficient accuracy. It can be observed in FIG. 4 that it is not trivial to find a low-degree polynomial that fits the variations in the path gain at different distances from the transmitter.

[0061] Ray-tracing offers high-fidelity channel modeling but remains computationally expensive, especially when creating the large-scale datasets needed for training DL models. To enable scalable dataset generation, we explore the trade-off between simulation accuracy and runtime by varying raytracing configurations.

[0062] Among various ray-tracing tools, Sionna™ RT (version 0.19.2) was selected for its ease of automation, native Python interface, and GPU acceleration — features that streamlinelarge-scale dataset generation. Sionna™ RT offers a flexible and efficient framework for radio propagation modeling, making it particularly well-suited for Al-driven wireless system design.

[0063] The area shown in FIG. 5 was used as a representative urban environment and 100 randomly selected TX locations were placed outside buildings. FIG. 5 is a geographic study area in Boston, MA, in which the red-outlined region marks the environment used for dataset generation, while the blue-highlighted region corresponds to the area used in the measurement campaign shown in FIG. 1. For each ray-tracing configuration, path gain radio maps are generated, and accuracy is evaluated using the RMSE metric. The configuration with diffraction enabled and a maximum depth of 20 is used as the ground truth (maximum configuration). All other configurations are compared against this baseline to identify the most efficient setup that balances accuracy and simulation time. All simulations are conducted using Sionna RT on an NVIDIA L40S GPU. Reported runtimes exclude scene initialization and geometry loading to focus solely on ray-tracing execution.

[0064] FIG. 6 shows the simulation runtime for 100 scenarios (each repeated ten times, with shaded 95% confidence intervals) across different ray-tracing configurations. As expected, enabling diffraction and increasing the maximum path depth both lead to longer runtimes. However, when diffraction is enabled, the RMSE levels off beyond a path depth of 10, indicating diminishing returns in accuracy. FIG. 7A illustrates how varying these parameters affects the resulting path-gain radio maps in a representative scenario. Based on this analysis, a “sweet spot” configuration is identified - diffraction enabled with a path depth of 10 - that balances fidelity and efficiency, requiring approximately 30 seconds to generate each radio map.

[0065] For all TX locations within the Boston area shown in FIG. 5, the scene geometry from the BostonTwin model (a digital twin for ray-tracing in 6G networks) is loaded into Sionna RT with a carrier frequency of 1 GHz. Both transmitter and receiver use vertically polarized isotropic antennas, and the transmit power is fixed at 44 dBm to match typical urban macrocell deployments. Path-gain radio maps are then generated for each TX position.

[0066] For training data, 12,000 valid TX locations are selected within the rectangle defined by latitudes 42.2796° to 42.3599° N and longitudes 71.1478° to 71.0453°W, excluding points inside building footprints. Each sample comprises a pair of uniformly cropped, rasterized maps: (1) the path-gain radio map and (2) the corresponding 2D building elevation map, with the TX centered in both. To introduce variability in coverage area, themap extent for each sample is drawn uniformly between 500m and 3 km. This pipeline can ingest any 3D urban dataset for which detailed geometry is available, enabling seamless generalization to new locations.

[0067] To mitigate overfitting and enhance model robustness, geometric data augmentation is applied: each sample is rotated by 90°, 180°, and 270°, and flipped horizontally and vertically. This expands the dataset to 60,000 samples, offering ample diversity for training deep-learning models. This represents a substantial radio-map dataset, with effectively unlimited extensibility in both geographic scope and sample count.Relevant Work on AI / ML-based Channel Models

[0068] In AI / ML-based channel modeling, a mapping function is established between the wireless environment and its corresponding channel properties, enabling accurate and realistic channel parameter representation. Over the past few years, numerous studies have explored and developed AI / ML-based channel models.

[0069] Early studies primarily employed conventional machine learning techniques, such as Random Forest, K-Nearest Neighbor (KNN), and Support Vector Machine (SVM) for channel path loss prediction, with Random Forest generally achieving the best performance. Seretis et al. later compared Random Forest with XGBoost, integrating extensive feature engineering to reduce model complexity. Their results indicated that XGBoost yielded higher prediction accuracy. A detailed summary of these works is presented in FIG. 7B.

[0070] Subsequent efforts shifted towards shallow neural network approaches. For instance, investigated techniques such as Radial Basis Function (RBF) neural networks to predict time-varying path loss, shadow fading, and small-scale channel characteristics, including the number of propagation paths and angular statistics, particularly within Geometry-Based Stochastic Model (GBSM). Huang et al. compared the performance of RBF networks with Multilayer Perceptron (MLP) neural networks, reporting that MLP slightly outperformed RBF by a fraction of a decibel.

[0071] Further, Zhang et al. utilized an MLP neural network with additional environmental features, including terrain type and building occlusion, to predict the signal strength coverage within an urban scenario. Their model demonstrated superior accuracy compared to traditional propagation models.

[0072] The introduction of Convolutional Neural Networks (CNNs) and their ability to learn spatial correlations in images led to a new generation of models that utilize images as input to neural networks. These approaches involve constructing images from propagation-related features, which CNNs analyze to extract spatial patterns, allowing the model to perform automated feature engineering and improve prediction accuracy.

[0073] Examples of image-based environmental representations include distance maps and building maps, as proposed by Imai et al., as well as low-resolution building height information. These CNN models are typically followed by fully connected (FC) layers for regression tasks, predicting channel parameters. Additionally, some studies have incorporated system parameters - such as frequency and antenna tilt - into the FC layers as extra features, as these parameters may not be easily represented within an image format.

[0074] Beyond structured maps, satellite and aerial images have also been integrated as additional input channels to CNN models. The impact of various environmental map images have been explored, while other works have examined how different image sizes and spatial map construction methods between TX and RX affect prediction accuracy. Their findings indicate that a building occupancy rate map is more effective than aerial imagery and that including both RX and TX image data achieves similar performance to adding system parameters.

[0075] Hybrid approaches have also been explored, where a physics-based model is used to generate rough parameter estimates, which are then refined using a neural network. This correction-based method improves prediction performance by leveraging domain knowledge in combination with deep learning. Rough estimates are integrated into the FC regression layer. Alternatively, a heat map of rough estimates - constructed using a free-space model -may be incorporated as an additional input channel for the CNN model.

[0076] Recent studies have focused on leveraging richer environmental information as input images to predict channel parameter heat maps in a single step. Other works have explored CNN-based autoencoders and U-Net architectures for fast and accurate predictions. While U-Net enhances prediction accuracy, it comes at the cost of increased computational complexity. A summary of DL-based methods for channel parameter prediction is presented in FIG. 7C.

[0077] The described embodiments are directed to a DL framework for estimating radio maps from environmental context, comprising separate training and deployment phases, as shown in FIG. 8. During training, the model learns to generate a spatial map of a target channel parameter — such as path gain or RMS delay spread — for any specified TX location. Unlike a previous approach, which required two inputs (an initial rough radio-map estimate plus an elevation map) and incurred significant overhead, the described embodiments useonly a single-channel 2D elevation image encoding terrain and building heights. The architecture extends an autoencoder inspired by PMNet, with adjusted encoder and decoder convolutional layers to process this one-channel input in place of PMNet’ s original dual -input design. The network structure is shown in FIG. 9A, with architectural details provided in FIG. 9B.

[0078] Choosing an effective input space is critical for enabling the model to accurately learn channel characteristics. Prior work has explored a variety of input encoding strategies. For example, Lee et al. used two input images: a building height heat map and a one-hot encoded image indicating the TX location. In contrast, Bakirtzis et al. incorporated richer environmental descriptors such as conductivity, permittivity, relative distance, and free-space path loss maps, primarily for indoor scenarios.

[0079] In the described embodiments, the input space is defined as x = Iei, where lei represents a 2D elevation map derived from the 3D model of the scenario. This map encodes terrain elevation, building heights, and other physical obstructions. Although the model input has a fixed spatial dimension (e.g., 200 x 200 pixels), we vary the spatial resolution of each sample is varied - ranging from 2.5 m / pixel to 15 m / pixel - so that the corresponding physical area spans from 500m to 3km per side. This resolution adaptive approach enables the model to learn propagation characteristics over different deployment scales.

[0080] The elevation map for our scenario is depicted in FIG. 10. Through experimental analysis, we found that applying minmax normalization, which scales all values between 0 and 1, enhances model convergence. Additionally, we invert the elevation values so that taller buildings — being the primary obstructions in the coverage map — are assigned values closer to zero. This transformation ensures that the model learns the impact of major blockages more effectively. Ensuring that the model accurately interprets the TX location is crucial. A literature review identified two effective methods for encoding this information: (1) representing the TX location as a one-hot-encoded heat map, where a single pixel in the image corresponds to the TX position on the building map, and (2) centering the TX within the input image, ensuring that the building map is always aligned around the TX location. At least some of the described embodiments adopt the latter as described herein.

[0081] For model evaluation, the entire dataset is divided using 5-fold cross-validation. Within each fold, we follow a 70-15-15 split for training, validation, and testing, respectively. To prevent data leakage, we ensure that augmented versions of the same scenario do not appear in the other sets. As shown in FIG. 11, the median error percentage across all foldsand coverage maps remains around 4.43%, demonstrating the model’s high accuracy. The path gain values in the test dataset range from -150 dB to -50 dB, encompassing a wide variety of propagation conditions. Despite this large dynamic range, the model maintains consistent performance with minimal deviations across different test cases, highlighting its robustness in estimating radio maps with exceptionally low inference error in real-time. Additionally, FIG. 10 presents qualitative results showcasing the model’s performance across various scenarios, which were randomly selected from the Boston area. The proposed model consists of 37,601,537 parameters, reflecting its capacity to learn complex spatial patterns in radio maps. To evaluate computational efficiency, we measured the inference time over 1,000 runs on an NVIDIA L40S GPU, achieving an average of 4.2 milliseconds per run. This low latency performance makes the model well-suited for real-time or large-scale radio map estimation in wireless systems and digital twins.Model Calibration

[0082] It is important to note that the ray-tracing simulations used for training were conducted at a frequency of 1 GHz, employing an isotropic antenna pattern. Additionally, the environmental geometry, while detailed, lacks full accuracy compared to the real-world scenario. Furthermore, the material properties assigned in the simulations differ from actual environmental materials, introducing discrepancies between simulated and measured propagation data. These factors contribute to the simulation-to-reality gap that the described embodiments are intended to bridge.

[0083] To address these discrepancies and improve the real-world accuracy of the AI-driven model of the described embodiments, a calibration pipeline for fine-tuning is introduced. FIGs. 12A and 12B show an example embodiment of a calibration pipeline using transfer learning with a weighted loss to align simulation with measurement data. The model is first pretrained on simulated radio maps to learn general channel characteristics. It is then fine-tuned using a small set of measurements to adapt to real-world conditions while preserving its generalization capabilities.

[0084] Given the cost and complexity of large-scale data collection, only a 20% subset of measurements is used for model calibration. This enables effective fine-tuning while preserving an unbiased portion of the data for testing. A geographically aware train / test split that clusters points by spatial region is adopted. This approach is more appropriate for geographic data, as it enforces spatial separation between training and testing areas, thereby providing a more realistic evaluation of the model’s ability to generalize across space. Toincorporate sparse measurements into training, the loss function is further modified using a weighted formulation that emphasizes measured points without overriding the knowledge learned from simulation.

[0085] This strategy enables efficient fine-tuning with minimal data, significantly improving predictive performance. The calibration process is performed 100 times to account for randomness in the train-test splits, ensuring the results are not biased toward any specific subset of the data. FIG. 13 shows the performance comparison (using empirical cumulative distribution function (eCDF) of error) with and without calibration, between Sionna RT 1302, the calibrated U-Net model 1304 of the described embodiments, and the uncalibrated model 1306. As shown, the calibrated model consistently outperforms the others, reducing the error to around 10% and producing predictions that more closely align with real -world measurements. This calibration pipeline allows the U-Net model to continuously incorporate real-world measurements, ensuring that the digital twin remains closely aligned with the physical network. By iteratively collecting field data, updating the model, and regenerating the radio maps, the system maintains high fidelity and accuracy for subsequent “what-if’ analyses and control actions. Despite Sionna RT’s support for gradient-based calibration of material and antenna parameters, we did not apply this process to our comparisons for two main reasons. First, the calibration methods in Sionna RT rely on full Cox-Ingersoll-Ross (CIR) and differentiable parametrizations of scattering and antenna patterns, whereas the measurements of the described embodiments supply only scalar path-gain values. Without access to absolute delays, phases, or multi-tap CIR data, the Sionna calibration pipeline cannot be directly applied. Second, the end-to-end differentiable calibration in Sionna RT incurs substantial computational overhead — requiring repeated ray tracing and backpropagation through complex computational graphs — which contradicts our real-time objectives. In contrast, the lightweight U-Net calibration procedure of the described embodiments uses only sparse path-gain samples and executes in milliseconds, ensuring both agility and fidelity for digital-twin applications.

[0086] The ultimate objective of these radio maps is to enable real-time estimation of channel parameters across a geographic area. The next step is to integrate the U-Net model into higher-layer workflows for system-level evaluation. In this section, we demonstrate and evaluate how the U-Net serves as a channel-parameter generator for both Sionna SYS system simulations, and Colosseum - the world’s largest wireless network emulator.Evaluation on Sionna SYS

[0087] Sionna SYS provides a modular framework to simulate full-stack system-level behavior. In each time slot, the system performs user scheduling across the resource grid, allocates transmit power to each user, computes the Signal to Interference plus Noise Ratio (SINR), selects the Modulation and Coding Scheme (MCS) via link adaptation, and generates decoded bits along with Hybrid Automatic Repeat reQuest (HARQ) feedback using physical layer abstraction.

[0088] In this framework, perfect channel state information is assumed for precoder and equalizer design, achievable-rate estimation (for scheduling decisions), and channel-quality feedback (for link adaptation). Performance is evaluated using both Shannon capacity and Outer Loop Link Adaptation (OLLA)-adjusted spectral efficiency, where OLLA dynamically selects the optimal MCS level.

[0089] We evaluate how path gain prediction errors impact system level performance, focusing on spectral efficiency and Block Error Rate (BLER), as well as the feasibility of real-time simulation. For this analysis, the Sionna SYS simulator is used, which enables realistic wireless network simulations through a flexible and Python-based interface.

[0090] FIGs. 15A, 15B, and 15C compare errors for three channel-modeling approaches - Sionna RT 0.19.2, the uncalibrated U-Net, and the calibrated U-Net - against measurement data. FIGs. 15A, 15B, and 15C depict error distribution ECDFs comparing measurement data, Sionna RT, and both uncalibrated and calibrated UNet models. FIG. 15A shows the error distribution across Shannon capacity, FIG. 15B shows the error distribution across link adaptation efficiency, and FIG. 15C shows the error distribution across block-error rate. Each model employs a single-tap channel with fixed propagation delay and randomized phase to focus exclusively on path-gain effects in system-level evaluations.

[0091] The results show that, despite a 10% error in path-gain estimates, the calibrated U-Net of the described embodiments achieves virtually near-zero error across system-level metrics when compared to measurement-based channels. The uncalibrated U-Net also outperforms the ray-tracing baseline. These outcomes confirm that the U-Net’ s residual pathgain deviations do not impact higher-layer simulations, validating its applicability for realtime system-level wireless emulation.

[0092] We implemented the full pipeline — from elevation map to system-level metrics — using Sionna SYS on an NVIDIA L40S GPU. After warm-up, each step is processed in 0.75 ms, and full radio map generation (200x200 points) takes 280 ms (284 ms including radiomap processing), enabling real-time operation. The pipeline leverages TensorFlow XLA with mixed precision and batch parallelism, achieving 14,005 steps / sec / point for rapid adaptation in dynamic wireless environments. FIG. 16 shows Shannon capacity heatmaps for ray tracing, uncalibrated U-Net, measurement data, and calibrated U-Net model estimations on test measurement points.Deployment on Colosseum

[0093] To enable real-time channel emulation as well as simulation using our U-Net model, the model was integrated into the Colosseum platform. Colosseum was extended with real-time scenario generation capabilities. As shown in FIG. 14, when a user selects or updates a location, the corresponding building elevation map is loaded, transformed into the appropriate input shape and format within 20ms, and passed to the U-Net model. The model performs inference in approximately 4ms, producing a path gain prediction. This result is then sent to an MQTT broker, which forwards it to Colosseum’s MCHEM, the FPGA-based Massive Channel Emulator. MCHEM then uses these taps to emulate the corresponding wireless channel in real time. In FIG. 17, we show results from running OpenAirInterface (OAI) on the Colosseum testbed using single-tap time domain channels with simple propagation delays to simulate mobile scenarios across four channel models: Ray Tracing, Calibrated U-Net, Uncalibrated U-Net, and actual measurements. Each experiment was repeated 10 times. Both the Ray Tracing and Uncalibrated U-Net models failed to establish a connection under the same configuration, resulting in no Reference Signal Received Power (RSRP) data. In contrast, the Calibrated U-Net model successfully produced RSRP patterns closely matching empirical measurements, with an RMSE of 8.86 dBm.

[0094] While example embodiments have been particularly shown and described, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the scope of the embodiments encompassed by the appended claims.

Claims

CLAIMSWhat is claimed is:

1. A method of estimating or predicting a radio map consisting of a coverage map, or any radio network characteristic, from environmental context, comprising:designating a data-driven training area;for each transmitter location in the training area,(i) generating a channel parameter radio map that is centered on the transmitter and is based on a ray-tracing-twin or any other channel model;(ii) generating a two-dimensional elevation map, centered on the transmitter, corresponding to the training area;training a U-Net-based model or any other neural network-based model using the channel parameter radio maps and the two-dimensional building elevation maps corresponding to the transmitters in the area, to produce a trained U-Net-based or any other neural network-based model;designating an area to be radio-mapped;providing the trained U-Net-based or any other neural network-based model with a two-dimensional input elevation map derived from a three-dimensional model of the area to be radio mapped; andproducing, by the U-Net-based or any other neural network-based model, an estimate of the radio map based on the two-dimensional elevation map.

2. The method of claim 1, wherein the channel parameter is path gain.

3. The method of claim 1, wherein the channel parameter is delay spread.

4. The method of claim 1, further comprising calibrating the trained U-Net-based or any other neural network-based model using one or more channel characteristic field measurements.

5. The method of claim 1, further comprising calibrating the trained U-Net-based or any other neural network-based model using a subset of collected channel characteristic measurements.

6. The method of claim 4, further comprising modifying a loss function associated with calibrating the U-Net-based or any other neural network-based model using a weighted formulation that emphasizes measured points without overriding knowledge learned from simulation.

7. The method of claim 1, further comprising configuring a set of the two-dimensional input elevation maps to each have a fixed spatial dimension, wherein a spatial resolution from map to map varies within the set.

8. The method of claim 1, further comprising applying minmax normalization by the U- Net-based or any other neural network-based model.

9. The method of claim 1, further comprising inverting elevation values within the two- dimensional input elevation map processed by the U-Net-based or any other neural network-based model.

10. The method of claim 1, further comprising generating the two-dimensional input elevation map by encoding building heights, terrain elevation, and physical obstructions within the area to be radio mapped.

11. A system for estimating a radio map consisting of a coverage map, or any radio network characteristic from environmental context, comprising:a U-Net-based or any other neural network-based model trained using one or more channel parameter radio maps and one or more two-dimensional building elevation maps;a two-dimensional input elevation map derived from a three-dimensional model of an area to be radio mapped;the two-dimensional input elevation map applied to an input of the U-Net- based or any other neural network-based model produces an estimate of the radio map at an output of the U-Net-based or any other neural network-based model.

12. The system of claim 11, wherein the one or more channel parameter radio maps is generated based on a ray-tracing twin model, and wherein the two-dimensional elevation map is generated corresponding a training area.

13. The system of claim 12, wherein the U-Net-based or any other neural network-based model is trained with the channel parameter radio maps and the two-dimensional building elevation maps corresponding to the transmitters in the area.

14. The system of claim 11, wherein the channel parameter is path gain.

15. The system of claim 11, wherein the channel parameter is delay spread.

16. The system of claim 11, wherein the U-Net-based or any other neural network-based model is trained with a subset of collected channel characteristic measurements.

17. The system of claim 11, wherein the U-Net-based or any other neural network-based model is trained with a subset of collected channel characteristic measurements.

18. The system of claim 11, wherein the U-Net-based or any other neural network-based model uses minmax normalization.

19. The system of claim 11, wherein the two-dimensional input elevation map uses inverted elevation values.

20. A method of estimating a radio map from environmental context, comprising:designating a training area;generating a channel parameter radio map that is based on a ray-tracing twin model;generating two-dimensional elevation maps corresponding to the training area; training a deep learning model using the channel parameter radio maps and the two-dimensional building elevation maps to produce a trained U-Net-based or any other neural network-based model;designating an area to be radio-mapped;providing the trained U-Net-based or any other neural network-based model with a two-dimensional input elevation map derived from a three-dimensional model of the area to be radio mapped; andproducing, by the U-Net-based or any other neural network-based model, an estimate of the radio map based on the two-dimensional elevation map.

Images