Multi-Modal Conditioned High-Resolution Transformer for Urban Electromagnetic Field Map Prediction Download PDF

2026-06-25

AI Summary

This study addresses the high computational cost of high-resolution electromagnetic (EM) field prediction in urban environments, where traditional physics-based simulations are inefficient. The authors propose a multi-condition dense prediction framework that takes building layout images and antenna configurations as inputs to generate 500×500 EM field maps. Their approach innovatively integrates FiLM modulation with cross-attention mechanisms for multimodal condition injection, incorporates transmitter-relative spatial encoding to enable coordinate-consistent test-time augmentation, and employs a composite loss function—combining masked L1, MS-SSIM, and focal L1—to mitigate imbalanced prediction difficulty. Built upon an HRFormer backbone, the model achieves a test-set MAE of 0.0461, outperforming UNet and plain HRFormer baselines by 25.2% and 31.8%, respectively, with test-time augmentation further reducing MAE by 6.3%.

Abstract

Predicting electromagnetic field (EMF) strength in urban environments is essential for cellular network planning but computationally expensive with physics-based simulators. We propose a multi-conditioned dense prediction framework that generates 500 500 EMF maps from building layout images and antenna configurations. Our architecture uses a High-Resolution Transformer (HRFormer) backbone with two complementary conditioning mechanisms: Feature-wise Linear Modulation (FiLM) injects scalar antenna parameters into all backbone stages, while cross-attention fuses 1-D radiation pattern tokens with spatial features at the deepest stage. We further introduce transmitter-relative spatial channels encoding distance, proximity, and bearing from the antenna, enabling coordinate-consistent test-time augmentation (TTA) that reduces test MAE by 6.3%. To address the prediction difficulty imbalance across EMF maps, we design a composite loss combining masked L1, multi-scale structural similarity (MS-SSIM), and a focal L1 term that upweights high-signal pixels, outperforming individual loss components in all metrics. Our best model achieves a test MAE of 0.0461, a 25.2% improvement over a plain UNet baseline and 31.8% over an HRFormer-only baseline.Do-

Problem

• electromagnetic field prediction
• urban environment
• cellular network planning
• computational efficiency

Innovation

• High-Resolution Transformer
• Feature-wise Linear Modulation
• cross-attention conditioning
• transmitter-relative spatial encoding
• composite loss