You are here

Drone Swarm Detection Using Artificial Intelligence Based on Ultrafast Neural Networks


OUSD (R&E) CRITICAL TECHNOLOGY AREA(S): Advanced Computing and Software


OBJECTIVE: Drone swarm detection using Artificial Intelligence. Develop a neural network architecture, and learning processing algorithms for drone identification based on ultra-fast neural networks with low power consumption.


DESCRIPTION: The goal of this Army Small Business Technology Transfer (STTR) topic is the identification of airborne drones from their radio frequency transmissions or radar signatures using ultra-fast neural networks [1]. The drone identification and classification are done by rapidly analyzing RF signals from one or several receiving antennas. As several target drones are often present in the antenna’s range, the received signal may represent a result of interference of several sources. For such a multiple target identification in a drone swarm, it is crucial to be able to process information in parallel and directly at the carrier microwave frequency.


Recent research into applications of artificial intelligence (AI) based on neuromorphic networks (in particular, using magnetic artificial neurons [1-5]) was executed to solve a variety of computational and signal processing problems. The purpose of neuromorphic computing is to replicate the human brain functionality in nanoscale using man-made neurons and synapses. The advantage of this approach is highly parallelized computing with large amounts of memory in close proximity to the computing elements, which results in a substantially increased speed and reduced power consumption of computing.


The methodologies described in [1-5] are particularly suited for defense-related computing due to a number of unique features, such as nano-scale sizes, simple implementation of memory elements and strongly nonlinear dynamics. Of a particular interest for military applications is the low power consumption of the network elements, and possibility of operation in GHz [2, 3] and even THz [4, 5] frequency ranges. These high-frequency properties allow one to utilize neural networks for parallel processing of drone microwave signals at the carrier frequency without digitization or super-heterodyning.


Another important consideration in the drone identification problem is the power requirements of the device. Recently, it has been demonstrated [6], that neural networks based on artificial antiferromagnetic neurons are capable of performing simple identification tasks in sub-nanosecond time with extremely low power consumption of less than 1 pJ per synaptic operation. These results look very promising for the development of mobile ultra-fast and low-power devices for neuromorphic identification of drones.


The goal of this call is to develop a neural network capable of simultaneous ultra-fast (time scale of nanoseconds) identification and targeting of large groups (swarms) of drones threatening ground vehicles. Another goal is to design an optimal architecture of an ultra-fast neural networks with integrated memory, and to develop and test learning and data-processing network algorithms suitable for ultra-fast detection of multiple drones in a drone swarm.


PHASE I: Using computer simulations, demonstrate the possibility of using artificial intelligence in the form of an ultra-fast neural network for processing multiple microwave drone signals without super-heterodyning or/and digitization. Demonstrate possibility of classification of drone microwave signals using ultra-fast neural networks in a case when the input signals from drones are monochromatic (unmodulated).


PHASE II: Determine optimum materials for the development of ultra-fast, lower power consumption neural networks. Develop principles of building large neural networks that will utilize ultra-fast processing capabilities of the chosen network elements (artificial neurons). Develop and test learning algorithms for drone identification in the presence of a single and multiple (2-5) drone signatures and modulated drone signals. Using computer simulations, demonstrate successful drone classification using a developed ultra-fast neural network. Determine processing time, power consumption, weight and size of an anti-drone device based on neural networks.


PHASE III DUAL USE APPLICATIONS: Demonstrate successful drone identification using an experimental prototype of a developed neural network. Demonstrate possibility of simultaneous identification of multiple drone targets. Potential applications include: light weight, ultracompact antenna for use in reconnaissance and observation drones (commercial and military); real-time monitoring of frequency agile microwave K band signals with potential applications to Active and Passive protection systems. Commercial application: Autonomous driving platforms and radar-based collison avoidance systems.



  1. J. Grollier, D. Querlioz, K.Y. Camsari, et al., “Neuromorphic spintronics.” Nat. Electron. 3, 360–370 (2020).;
  2. A. Ross, N. Leroux, A. De Riz, et al., “Multilayer spintronic neural networks with radiofrequency connections.” Nat. Nanotechnol. (2023).;
  3. J. Torrejon, M. Riou, 3.  F. Araujo, et al., “Neuromorphic computing with nanoscale spintronic oscillators.” Nature 547, 428–431 (2017).;
  4. R. Khymyn, I. Lisenkov, J. Voorheis, et al., “Ultra-fast artificial neuron: generation of picosecond-duration spikes in a current-driven antiferromagnetic auto-oscillator.” Sci. Rep. 8, 15727 (2018).;
  5. H. Bradley, S. Louis, C. Trevillian, et al., “Artificial neurons based on antiferromagnetic auto-oscillators as a platform for neuromorphic computing.” AIP Advances 13, 015206 (2023).
  6. H. Bradley, S. Louis, V. Tiberkevich, et al., “Pattern recognition using spiking antiferromagnetic neurons.” (Aug 2023).


KEYWORDS: artificial intelligence, ultra-fast, artificial neuron, drone identification, learning algorithm

US Flag An Official Website of the United States Government