SPAICE - Satellite Signal Processing Techniques using a Commercial Off-The-Shelf AI Chipset

Status date
Activity Code

The SPAICE project aims to study, develop, and validate Artificial Intelligence (AI)-based signal processing techniques for satellite communications in scenarios and use cases where specific AI processors can provide a significant performance improvement with respect the current state-of-the-art.

The team has already identified and traded-off some prospective scenario candidates such as interference detection and mitigation, interference localization, link adaptation, forward error correction, channel estimation, flexible payloads, and adaptive beamforming; all of them targeted for regenerative satellites. In addition, the team has also investigated on the suitable ML architectures, frameworks and Commercial-Off-The-Shelf (COTS) AI chipsets that can bring greater benefits to their combined use in the abovementioned scenarios targeting onboard communications satellites. A combination of flexible resource allocation payload and adaptive beamforming was selected for the demonstration phase.

The SPAICE project has as its principal outcome the AI Satellite Telecommunications Testbed (AISTT), which will be the platform to test and demonstrate the selected AI-accelerated scenarios. The AISTT will take advantage of the experimental expertise and available facilities at the SnT―University of Luxembourg such as the CubeSat Laboratory and the Satellite Communications Laboratory.

 Once the AISTT is successfully designed, implemented and the selected AI-accelerated scenarios are successfully tested and validated, the last objective of SPAICE is the evaluation the potential road-to-the-market of the testbed.


The targeted improvement of this activity with respect to the existing state-of-the-art is enabling on-board real-time satellite signal processing techniques. Existing solutions usually exhibit inherent limitations in translating theory to practice when handling the computational complexity and/or the latency required for the outcomes specially when dealing with large search space or high degrees of freedom. This has been typically the motivation for the use of AI, since it has been shown to have a strong potential to overcome this challenge via data-driven solutions. In the SPAICE project, we evaluated what are the expected gains, both in terms of latency, complexity and performance that can be achieved with the application of AI-based techniques in each of the considered use cases.


After an initial trade-off of multiple scenarios, the SPAICE project focuses on the AI-acceleration of a combination of flexible resource allocation and adaptive beamforming payload on-board regenerative satellites.

Next generation of satellite communications are built with the capability to assign radio resources quickly and flexibly according to the system load and the changing environment. Many important resource allocation problems are non-convex or combinatorial in nature because of the discrete nature of the variables involved (e.g., carrier allocation, user scheduling). Hence, computing the optimal solution is very challenging, leading to unaffordable computational times.

Flexible EIRP over the service area

Furthermore, antenna arrays play an increasingly important role as we move into the era of high-capacity wireless access systems that demand high spectral efficiency. However, when talking about how to autonomously manage and adapt the antenna parameters to modify the radiation pattern, the number of available solutions decrease due to the high complexity of the algorithms.

Dynamic antenna beamforming

However, the SPAICE project proposes that resource management and adaptive beamforming can be performed onboard the satellite thanks to the use of ML techniques running in onboard AI accelerators to dynamically adjust the bandwidth, power and beamwidth of multi-beam regenerative satellites.


The SPAICE project proposes a Supervised Learning (SL) approach where the model is trained on-ground and the inference is executed on-board the satellite. The ML algorithms implemented as Neural Networks (NN) will be cascaded during the inference on-board the satellite.

Cascaded ML-based joint resource allocation and beamforming

A classification algorithm based on Convolutional NN (CNN) is proposed to implement the flexible payload algorithm. The pre-processing required in a CNN is much lower compared to other NN-based classification algorithms such. In that sense, the CNN can successfully capture the spatial and temporal dependencies of the input by applying the relevant filters. The CNN architecture is optimized compared to other classification algorithms to improve the processing performance of a data set due to the reduction of the number of parameters and the reuse of weights.

On the other hand, a Feedforward NN (FNN) implementation to approximate a beamformer that gives as output the beamforming weights. As an input layer of the FNN, we consider loading several features, that are processed by the neurons of the hidden layers. In particular, we consider training the NN to derive the weights that can produce the target radiation pattern taking into account the power and minimum side-lobe level constraints.

System Architecture

The system simulator was built using MATLAB and feeds the cascaded joint resource allocation and beamforming algorithm with realistic geographical-based traffic demand using the Satellite Traffic Emulator developed by the SnT SIGCOM group.

Block diagram describing the system architecture

A channel simulator, an antenna pattern simulator and a link budget simulator are used to compute the served throughput and compare it with the original geographical traffic demand.

The proposed ML algorithms are implemented using Tensorflow and trained using the system simulator data. For the inference part used in the AISTT, the initial AI accelerator trade-off has shown that the Xilinx Versal AI ACAP family is the best COTS chipset in terms of performance versus consumed power from their specifications.

Operations per second per Watt



The project has 2 phases with total duration of 26 months.

The WBS is composed of the following WPs:

  • WP1 Scenario and Requirements Definition (Phase 1)

  • WP2 ML Preparation and Generation of Data Sets (Phase 1)

  • WP3 ML Setup, Training and Validation (Phase 2)

  • WP4 AISTT Requirements, Architecture and Design (Phase 2)

  • WP5 AISTT Test Plan and Implementation (Phase 2)

  • WP6 AISTT Validation and User Manual (Phase 2)

  • WP7 Roadmap and Lessons Learned (Phase 2)

  • WP8 Project Management (Phase 1 & 2)

Current status

As of December 2022, the SPAICE project completed its Phase 1 with a successful definition of the scenarios and requirements, definition of the ML algorithms and their framework, generation of the training data sets, and selection of the AI chipset family.

The SPAICE team is now working towards the training of the ML algorithms and the definition of the AISTT in the Phase 2 of the project.

Prime Contractor