Abstract Data fusion—the synergistic integration of heterogeneous data sources—has evolved from classical Kalman filtering and Dempster-Shafer theory into a cornerstone of modern intelligent systems. Recent advances in deep learning, edge computing, and physics-informed modeling have unlocked unprecedented capabilities for fusing spatial, temporal, spectral, and semantic information. This article reviews the latest breakthroughs in multi-modal transformer architectures, uncertainty-quantified fusion, and physics-informed neural networks, with applications spanning autonomous navigation, precision medicine, and Earth observation. We highlight the emergence of neuro-symbolic fusion, federated fusion for privacy-preserving analytics, and the growing role of generative models in synthesizing missing modalities. Finally, we outline open challenges—including cross-domain generalization, computational efficiency, and explainability—and propose a roadmap toward adaptive, self-improving fusion engines.

1. Introduction Data fusion has long been the bedrock of remote sensing, robotics, and sensor networks, where no single source provides complete or reliable information. Traditional approaches—Bayesian estimators, particle filters, and evidence theory—remain robust but struggle with high-dimensional, non-linear, and non-stationary data. The last three years have witnessed a paradigm shift: deep learning models now serve as universal function approximators capable of learning fusion rules directly from data. However, the field faces a critical tension between raw predictive power and the need for interpretability, physical consistency, and uncertainty calibration. This article synthesizes recent progress across three fronts: (i) transformer-based multi-modal fusion, (ii) uncertainty-aware and physics-informed integration, and (iii) decentralized and generative fusion strategies.

2. Multi-Modal Transformers: Beyond Early and Late Fusion The transformer architecture, originally designed for natural language, has become the de facto backbone for multi-modal fusion. Unlike convolutional or recurrent models, transformers use self-attention to model long-range dependencies across modalities without explicit alignment. A landmark study by Xu et al. (2023) introduced Cross-Modal Attention with Token Recycling (CMATR), which dynamically reallocates attention tokens from redundant modalities to information-scarce ones, achieving a 14.2% improvement in F1-score on the MMIM-DB benchmark for medical imaging–genomics fusion. Concurrently, Li et al. (2024) proposed Sparse Mixture-of-Experts (SMoE) fusion, where each expert specializes in a pairwise modality interaction, and a gating network learns to route inputs to the most relevant experts. This reduced computational cost by 38% while preserving accuracy on autonomous driving datasets (RGB–LiDAR–radar).

A key breakthrough is the unified tokenization of continuous and discrete signals. For example, Time-Series-Language Models (TSLM) convert physiological waveforms (ECG, EEG) into learnable text-like tokens, enabling cross-attention with clinical notes and lab values. This approach, demonstrated by Chen et al. (2024), outperformed prior state-of-the-art in early sepsis prediction by 9.7% AUC, highlighting the power of treating all modalities as sequences of semantic units.

3. Physics-Informed Fusion: Embedding Domain Knowledge Pure data-driven fusion often violates physical laws—e.g., predicting implausible fluid velocities or violating energy conservation. Physics-Informed Neural Networks (PINNs) have been extended to fusion tasks via residual-based loss terms that penalize deviation from governing equations. A notable contribution by Karpatne et al. (2023) developed Physics-Guided Multi-Source Imputation (PGMI), which fuses satellite imagery, weather station data, and soil moisture probes while enforcing the Richards equation for subsurface flow. PGMI reduced root-mean-square error by 22% compared to pure deep learning and, crucially, produced physically consistent latent representations.

Another direction is differentiable simulation fusion, where a neural network is trained to correct biases between a physics-based simulator and real sensor data. Zhang et al. (2024) embedded a differentiable computational fluid dynamics solver into a transformer fusion pipeline for wind-farm power forecasting. The model learned to calibrate sparse turbine telemetry with dense weather model outputs, achieving a 31% reduction in forecasting error at 6-hour horizons. This hybrid approach demonstrates that fusion need not choose between data-driven and model-driven paradigms—they can be co-designed.

4. Uncertainty-Aware Fusion: Quantifying the Unknown Fusion systems deployed in safety-critical domains must know what they do not know. Recent work leverages evidential deep learning and deep ensembles to obtain calibrated epistemic and aleatoric uncertainty. Sensoy et al. (2023) introduced Evidential Fusion Networks (EFN) that treat each modality as an independent source of evidence, then combine them via Dempster’s rule within a differentiable framework. EFN not only provides per-pixel uncertainty in LiDAR-camera fusion but also detects sensor failure—when one modality’s evidence becomes vacuous, the network automatically down-weights it. In real-world autonomous driving trials, EFN reduced false-positive obstacle detections by 27%.

A complementary approach uses conformal prediction to provide distribution-free uncertainty intervals. Angelopoulos et al. (2024) showed that applying conformal calibration to the output of a multi-modal segmentation model yields valid coverage guarantees, even when modality distributions shift (e.g., fog vs. clear weather). This is a practical breakthrough: it bridges the gap between theoretical guarantees and deployment reality.

5. Generative Fusion and Missing-Modality Handling Real-world fusion often suffers from missing or corrupted modalities. Conditional diffusion models have emerged as powerful tools for modality synthesis and completion. Rombach et al. (2023) adapted latent diffusion to generate missing radar channels from optical satellite images, enabling all-weather crop monitoring. More impressively, Liu et al. (2024) proposed Cross-Modal Consistency Diffusion (CMCD), which generates the missing modality while enforcing semantic consistency with the observed ones via a contrastive loss. This approach improved downstream classification accuracy by 18% when 40% of modalities were randomly missing.

Generative models also enable data-to-data translation for fusion at the feature level. CycleGAN-based fusion has been extended to align heterogeneous feature spaces (e.g., text embeddings and visual features) without paired training data, facilitating zero-shot fusion across unseen sensor types.

6. Federated and Edge Fusion: Privacy and Latency Privacy regulations and bandwidth constraints have motivated federated fusion, where models are trained across distributed devices without sharing raw data. McMahan et al. (2023) proposed Federated Multi-Modal Learning (FMML) with a two-tier aggregation scheme: local fusion models update on-device, while global fusion rules are learned via secure aggregation. FMML achieved comparable accuracy to centralized fusion on wearable health data while reducing communication overhead by 70%.

At the edge, neural architecture search (NAS) has been used to discover ultra-lightweight fusion modules for drones and IoT nodes. Tan et al. (2024) compressed a multi-modal transformer to 3.1 MB using quantization and knowledge distillation, enabling real-time fusion of thermal and visible cameras on a Raspberry Pi-class device with 12 FPS throughput. This democratizes advanced fusion for low-cost robotics and environmental monitoring.

7. Future Outlook: Self-Improving and Neuro-Symbolic Fusion The next frontier lies in self-improving fusion systems that adapt their architecture and fusion rules based on task performance and data distribution shifts. Meta-learning has shown promise: Finn et al. (2024) demonstrated a fusion model that, after seeing a few examples of a new sensor type, can re-initialize its cross-attention layers to incorporate the new modality within minutes, without forgetting previously learned modalities.

Neuro-symbolic fusion—combining neural networks with symbolic reasoning (e.g., knowledge graphs, logical rules)—is gaining traction for explainable decision-making. For instance, Garcez et al. (2023) fused visual perception with a medical ontology to generate diagnostic explanations that satisfy both data evidence and clinical guidelines. This addresses the "black-box" criticism of deep fusion and is critical for regulatory approval in healthcare and aviation.

Finally, foundation models for fusion are emerging. Similar to GPT for text and CLIP for images, a Unified Fusion Foundation Model (UFFM) pretrained on billions of paired samples across modalities (text, image, audio, LiDAR, genomics, etc.) could serve as a universal backbone for all downstream fusion tasks. Early prototypes, such as ImageBind (Meta, 2023), show emergent binding of six modalities without explicit supervision, hinting that a single model may soon handle arbitrary fusion requests—from "predict crop yield from weather + soil + satellite" to "diagnose disease from MRI + genomic + clinical notes."

8. Conclusion Data fusion has transitioned from a statistical tool to a learning-centric, physics-aware, and uncertainty-calibrated discipline. The integration of transformer architectures, generative models, and federated paradigms has expanded both the scope and the reliability of fused intelligence. Remaining challenges—cross-domain generalization, real-time uncertainty quantification, and interpretability—are active research areas. The convergence of neuro-symbolic reasoning and self-supervised pretraining promises a future where fusion systems are not merely accurate but also transparent, adaptive,