v11labs

Inverse Parameter Detection of RF Exposure Using Convolutional Neural Networks Trained on FDTD-Derived SAR Profiles

Abstract

Exposure to radiofrequency (RF) electromagnetic fields produces spatially structured energy deposition in biological tissue, commonly quantified through specific absorption rate (SAR). While forward modeling of SAR given known exposure conditions is well established, the inverse problem, inferring RF exposure parameters from observed SAR distributions, remains largely unexplored.

In this work, we present a machine-learning framework for inverse RF exposure reconstruction, using convolutional neural networks (CNNs) trained on SAR profiles generated via finite-difference time-domain (FDTD) simulations. Synthetic SAR datasets were produced across a wide range of RF frequencies, power densities, and angles of incidence using high-resolution voxelated human head phantoms. The trained models learn a direct mapping from post-exposure SAR “afterimages” to their originating RF parameters.

Results demonstrate that CNNs can reliably recover frequency and power density with low error across broad parameter ranges, while angular inference improves with targeted data augmentation. These findings establish the feasibility of AI-driven inverse RF exposure detection and support its application to exposure diagnostics, retrospective analysis, and real-time RF monitoring systems.

1. Introduction

Unexplained RF related physiological incidents have prompted renewed interest in understanding how complex exposure conditions manifest within biological tissue. While experimental and computational studies have extensively characterized forward RF tissue interactions, real world investigations are often hindered by the absence of reliable exposure metadata such as frequency, power density, and angle of incidence. This lack of information renders direct comparison across cases impractical and limits mechanistic interpretation.

Rather than attempting to infer exposure conditions indirectly from symptoms or injury patterns, we reformulate the problem as a well-posed inverse task: can RF exposure parameters be recovered directly from the spatial distribution of absorbed energy in tissue?

To address this question, we combine electromagnetic simulation with supervised deep learning. Using FDTD solutions to Maxwell’s equations, we generate large volumes of SAR data under controlled RF conditions and train CNN-based models to infer the originating exposure parameters. This approach leverages the fact that SAR encodes spatial signatures of frequency dependent penetration depth, power scaling, and angular coupling—features that are difficult to disentangle analytically but well suited to data-driven inference.

2. Methodology

2.1 Dataset Generation

Synthetic SAR datasets were generated using high-resolution FDTD simulations of RF propagation through a voxelated human head phantom. Models were constructed with isotropic voxel resolutions of 3 mm, 2 mm, and 1 mm, incorporating heterogeneous tissue dielectric properties to reflect realistic anatomical structure.

RF exposure parameters spanned the following ranges: Frequency: 400 MHz – 3000 MHz; Power Density: 0.1 – 1000 mW/cm² ; Angle of Incidence: 0° – 360°

For each configuration, simulations produced 10 g-averaged SAR volumes. Parameter values were randomly sampled and uniformly distributed to ensure broad coverage of the RF exposure space.

2.2 Preprocessing and Normalization

SAR volumes were normalized using a sigmoid transformation to bound dynamic range and stabilize training. Target parameters were similarly scaled to a bounded domain. Models were trained using mean squared error (MSE) loss.

Datasets were split 70/30 into training and testing sets, with the training subset further divided into training and validation partitions (70/30).

2.3 Model Architecture and Training

Models were implemented in PyTorch using feedforward neural networks operating on flattened SAR inputs. Architectures ranged from two to four hidden layers, employing progressive dimensionality reduction with ReLU activations. Output layers predicted three continuous values corresponding to frequency, power density, and angle. Training used adam's optimization with a learning rate of 0.001, 50-200 epochs, batch sizes of 16-64 and Low-Rank Adaptation (LoRA) was used to fine-tune higher-resolution models from pretrained lower-resolution foundation networks, enabling efficient transfer learning while preserving learned spatial representations.

3. Results

3.1 Foundation Models

Model 1.0 (3,778 samples)

The initial foundation model demonstrated clear learning behavior across all three parameters, with particularly strong performance in frequency and power density inference. Angular prediction error was higher, attributable to uneven angular sampling concentrated below 180°.

Despite limited dataset size and shallow architecture, the model achieved stable convergence with minimal overfitting, confirming that SAR distributions contain sufficient information to support inverse RF parameter inference.

Model 2.0 (5,778 samples)

Model 2.0 incorporated an additional 2,000 simulations with expanded angular coverage, directly addressing limitations observed in Model 1.0. Increased epochs and batch size further improved learning stability.

This model achieved a 72% reduction in final training loss relative to Model 1.0, with RMSE reductions of approximately 50% for power density and 60% for frequency. Angular prediction accuracy improved substantially, demonstrating the effectiveness of targeted data augmentation.

Model 3.0 (9,778 samples)

Model 3.0 expanded the power density range to 0.1–1000 mW/cm² and introduced a deeper architecture with batch normalization and dropout. Pretrained weights from Model 2.0 were used as an initialization point.

While power density inference improved achieving a 93% reduction in final loss frequency and angle accuracy degraded. This behavior is consistent with catastrophic forgetting, where fine-tuning on a skewed data distribution biases learning toward newly dominant features.

Additionally, reduced accuracy at frequency extremes is attributed to FDTD resolution limits: at higher frequencies, the 3 mm voxel grid approaches a significant fraction of the RF wavelength, introducing numerical dispersion and noise into the training data.

4. Discussion

4.1 Inverse RF Exposure Inference

This study demonstrates that CNNs can successfully learn the inverse mapping from SAR distributions to RF exposure parameters. Even relatively shallow architectures extract meaningful spatial features that encode frequency-dependent penetration, power scaling, and angular coupling effects.

These results position SAR-based inverse inference as a viable tool for RF exposure reconstruction, extending beyond traditional forward-only modeling approaches.

4.2 Safety, Policy, and Diagnostics

Inverse SAR inference enables retrospective exposure analysis in scenarios where direct RF measurements are unavailable. This capability has immediate relevance for regulatory compliance, forensic investigations, and RF safety assessment across military, industrial, and consumer domains.

Coupling SAR inference with bioheat modeling further enables prediction of thermal tissue response, supporting risk assessment for sensitive anatomical structures.

4.3 Toward Individualized and Real-Time Systems

By incorporating shielding simulations, individualized anatomical models, and biological variability, this framework can be extended to personalized exposure assessment and evaluation of protective equipment. The low inference cost of trained models also supports deployment in real-time, edge-based RF monitoring systems.

5. Future Work

Ongoing work includes higher-resolution simulations, integration of time-dependent bioheat models, expansion to additional RF parameters (polarization, modulation, pulse structure), and experimental validation using physical phantoms and controlled exposure studies.

Long-term, this approach enables a closed-loop RF diagnostics ecosystem combining sensing, inverse inference, biological modeling, and automated mitigation.