Synthefy: Multi-Modal Diffusion Models for Synthetic RF Time Series

Synthefy is developing advanced generative AI models to create synthetic radiofrequency (RF) time series data for defense applications, including electronic warfare and signal intelligence. Our platform generates realistic RF signals using diffusion models that mimic real-world environments, enabling stress-testing, augmenting under-sampled datasets, and training machine learning classifiers. A key innovation is our “time series CLIP” model, which learns a joint representation of RF signals and key contextual metadata (e.g., signal type, modulation, location, RF parameters) through contrastive learning on large public datasets.  Our “time series CLIP” model is the backbone of our novel diffusion model, which generates realistic RF signals on public NIST Radar and Tektronix datasets that significantly beat today’s state-of-the-art generative adversarial networks (GANs). Moreover, a byproduct of our “time series CLIP” model is that it learns a joint representation of RF time series and metadata. As such, it enables efficient storage and retrieval of vast RF data sets, allowing DoD to search and identify relevant signals quickly. For example, a signal intelligence engineer could ask: “Generate a synthetic radar jamming pattern observed in a recent conflict,” or “Find me all RF signals similar to this jamming pattern for further analysis.” These use cases showcase Synthefy’s ability to generate RF signals that replicate real-world conditions and enable fast semantic search through a vector database of stored RF time series data. By integrating contextual metadata with time series data, we generate more accurate synthetic signals than traditional methods, which is critical for defense applications where environmental factors impact signal characteristics. Deliverables of this project include functional AI models, code transfer, and technical reports showcasing model performance and capabilities.   Synthefy has raised over $6 million from top Silicon Valley VCs and is well equipped to handle a DoD project. As a testament to our technology, our company won Deutsche Telekom (DT) and T-Mobile's AI in Telecom Challenge, beating 360 companies including Amazon to win a cash prize of 150K Euros. We also are working with Dell, Samsung, AT&T, DT, Tektronix, Opanga Networks, among others.