2.1 Aircraft Health Monitoring and predictive maintenance

In today’s competitive airline industry, cost-consciousness and operational efficiency are paramount, driving the need for innovative approaches to cost reduction, particularly in maintenance. Despite recent improvements, unplanned maintenance still accounts for over 25% of maintenance spending and contributes to 5% of wasted fuel consumption. Predictive maintenance, coupled with data analytics, presents a promising solution to address these challenges, although implementation hurdles still exist.

Predictive maintenance utilises aircraft-generated and operational data to assess onboard systems’ health, monitored by sensors tracking key parameters. This data enables proactive maintenance scheduling, potentially preventing significant performance declines or system failures. Flight Data Acquisition Units (FDAUs) collect and analyse this data, with some information transmitted from avionics systems via data buses like ARINC-429. Various methods exist for accessing and transmitting predictive maintenance data, including real-time transmission via ACARS or WiFi, or storage on removable media like compact flash drives. However, challenges such as cost and data volume must be considered. Data cleaning is essential before analysis, involving the consolidation and error correction of diverse data formats and sources, including paper documents.

Recent years have witnessed the integration of AI and ML in aircraft maintenance, enabling anomaly detection and proactive component maintenance planning. However, challenges persist, such as the ‘no fault found’ (NFF) dilemma, where components exhibiting anomalous behaviour may pass testing, complicating maintenance optimisation. Modern aircraft generate vast amounts of data, posing challenges in data utilisation. While data collection is straightforward, establishing baseline conditions for anomaly detection is crucial. AI and ML algorithms play a pivotal role in identifying trends and exceedances, enabling proactive maintenance and operational practices as shown in Fig. 2.

Figure 2. Application of AI in predictive maintenance [Reference Doherty29].

In the aerospace domain, both aircraft and spacecraft require high levels of precision and safety. The adoption of AI represents a transformative strategy for addressing maintenance challenges. Moving away from traditional methods, AI utilises advanced sensing technologies, combined with ML and DL algorithms, to predict and mitigate issues before they become critical problems. Spacecraft, often equipped with extensive systems such as antennas, booms and solar arrays as shown in Fig. 3, are susceptible to the effects of transient thermal states and material fatigue, impacting their overall integrity and functionality. Traditional damage identification methods typically involve comparisons with undamaged counterparts, focusing on properties such as stiffness and mass. However, these methods struggle to detect minor damage. Addressing this limitation, Paolo et al. [Reference Iannelli, Angeletti, Gasbarri, Panella and Rosato30] introduced a Bi-LSTM network to classify structural damages based on the number of beam elements that have failed within a truss system.

Figure 3. Spacecraft with large and foldable structures [Reference Iannelli, Angeletti, Gasbarri, Panella and Rosato30].

Forecasting the RUL of spacecraft components is critical to minimising unnecessary maintenance and preventing unforeseen equipment failures. Danyang et al. [Reference Han, Yu, Gong, Song and Tian31] have advanced this domain by devising a convolutional recurrent neural network (CRNN) shown in Fig. 4 that uses stator current data to predict the RUL of spacecraft bearings. Additionally, Gang et al. [Reference Gang, Ran, Peng, Tian and Qu32] introduced a fault detection framework using spacecraft sensor data. Their model integrates a variational autoencoder (VAE) and a gated recurrent unit (GRU) network to effectively extract features from the data, facilitating real-time fault detection in spacecraft.

Figure 4. Block diagram of CRNN for RUL [Reference Han, Yu, Gong, Song and Tian31].

Sara et al. [Reference Ibrahim, Ahmed, Zeidan and Ziedan33] employed telemetry parameters to assess performance using a support vector machine for regression (SVR). Subsequent analyses, including k-means clustering, t-distributed stochastic neighbour embedding (t-SNE) and logical analysis of data (LAD), were utilised to construct the dataset. Following dataset creation, fault tree analysis was applied for fault detection in data derived from Egyptsat-1 shown in Fig. 5.

Figure 5. Egyptsat-1 [Credit: Egypt’s NARSS].

Moving on to aircraft systems, engines represent a critical component where predictive maintenance and fault detection methodologies play a pivotal role in enhancing fuel efficiency, optimising maintenance schedules and improving decision-making processes. Changchang et al. [Reference Che, Wang, Fu and Ni34] employed deep belief networks (DBN) for condition assessment and fault detection, followed by long short-term memory (LSTM) networks for RUL prediction. Their methodology, applied to the NASA C-MAPSS dataset, demonstrated lower error rates compared to traditional methods. Maria et al. [Reference De Giorgi, Campilongo and Ficarella35] developed a numerical tool that leverages artificial neural networks (ANN) and support vector machines (SVM) to simulate and predict engine performance under both healthy and degraded conditions. Their approach focuses on identifying faulty components by analysing engine data, which includes parameters such as altitude, Mach number and rotation speed, alongside measurable variables like exhaust gas temperature and fuel mass flow rate. This method facilitates a detailed assessment of engine health and performance anomalies. Like in spacecraft to predict RUL, Ade et al. [Reference Hermawan, Kim and Lee36] implemented a convolutional long short-term memory (ConvLSTM) network. This network was tailored to analyse data from 21 sensors within an engine shown in Fig. 6, utilising the NASA C-MAPSS dataset. Further, the authors extended their research by applying advanced deep learning techniques, specifically, temporal convolutional neural networks (TCNN) and transformers to the same NASA dataset, which encompasses readings from 26 sensors alongside three operational settings.

Figure 6. Engine structure used for NASA C-MAPSS dataset [Reference Liu, Zhang, Liao, Wu and Peng37].

The majority of the methodologies reviewed herein primarily focus on static analysis. However, Chuang et al. [Reference Chen, Zhu, Shi, Lu and Jiang38] introduced a dynamic predictive model shown in Fig. 7, distinct in its approach to forecasting health state (HS) and RUL. This model leverages an ensemble of deep autoencoders (DAE) paired with bidirectional long short-term memory networks (Bi-LSTM), and it has been trained utilising the NASA C-MAPSS dataset.

Figure 7. Block flow diagram of using DL to predict RLU [Reference Chen, Zhu, Shi, Lu and Jiang38].

In another application, AI is being utilised for automated aircraft maintenance inspection to detect structural dents shown in Fig. 8 on various aircraft components. While numerous defects are visually detectable, smaller defects might elude unaided human observation. Dogru et al. [Reference Doğru, Bouarfa, Arizar and Aydoğan39] addressed this issue by proposing a masked R-CNN framework that incorporates transfer learning. Their dataset was specifically designed to include a variety of dent sizes.

Figure 8. Dents of various size on aircraft structures [Reference Doğru, Bouarfa, Arizar and Aydoğan39].

2.2 Air traffic management

AI is revolutionising air traffic management (ATM) through diverse applications: ATM is the complex choreography that maintains safe and efficient air travel. As airspace becomes increasingly congested, researchers are employing AI to revolutionise ATM, envisioning a future where human expertise and machine intelligence work collaboratively. Brittain and Wei [Reference Brittain and Wei40] proposed a framework using deep multi-agent reinforcement learning for autonomous air traffic control systems. Trained in the BlueSky simulation environment, AI agents prioritise safety and efficiency, resolving conflicts in high-density traffic scenarios. In safety-critical domains like ATM, transparency is paramount, driving the adoption of explainable AI (XAI) frameworks. Xie et al. [Reference Xie, Pongsakornsathien, Gardi and Sabatini20] emphasise interpretable AI models within ATM (Fig. 9), investigating post-hoc explanations to elucidate decision processes. Augustin et al. [Reference Degas, Islam, Hurter, Barua, Rahman and Poudel41] extend this work with the DPP (descriptive, predictive, prescriptive) model (Fig. 10), enhancing human-centred AI systems in ATM.

Figure 9. Framework of prediction model [Reference Xie, Pongsakornsathien, Gardi and Sabatini20].

Figure 10. Synthesis of explainable AI (XAI) conceptual framework [Reference Degas, Islam, Hurter, Barua, Rahman and Poudel41].

Building on predictive applications, ML techniques are transforming ATM operations: the applications of AI in ATM extend beyond conventional air traffic control, notably exploring ML techniques in predicting ground delay programmes (GDPs) [Reference Liu, Liu, Hansen, Pozdnukhov and Zhang42]. These programmes, activated due to disruptions like adverse weather, leverage ML algorithms (SVMs, LR, RF) to process historical data, deriving a ‘regional convective weather variable’ that enhances GDP forecasting accuracy. Tang et al. [Reference Tang, Liu and Pan43] systematise these advances, highlighting AI’s role in automating tasks across ATM sectors (ATS, AM, ATFM, FO) while identifying data quality and user expertise as persistent challenges.

In operational contexts, AI addresses both air and space traffic challenges: in space traffic management, Vasile et al. [Reference Vasile, Rodrguez-Fernández, Serra, Camacho and Riccardi44] integrate machine learning with orbital mechanics for collision risk mitigation. For air traffic, Kistan et al. [Reference Kistan, Gardi and Sabatini45] demonstrate deep neural networks (DNNs) in Airborne Collision Avoidance System (ACAS) Xa, achieving 40% performance gains over Traffic Collision Avoidance System (TCAS), while Gallego et al. [Reference Gallego, Comendador, Nieto, Imaz and Valdés46] combine DBSCAN, multi-level models (MLM) and ANNs to analyse aircraft vertical trajectories and predict altitude levels. Evolutionary algorithms also show promise – Lertworawanich et al. [Reference Lertworawanich, Pongsakornsathien, Xie, Gardi and Sabatini47] apply genetic algorithms (GA) with K-means clustering to optimise direct capacity balancing (DCB) overload management.

ML-driven traffic flow analysis exemplifies AI’s spatial-temporal capabilities: Gui et al. [Reference Gui, Zhou, Wang, Liu and Sun48] analyse air traffic flow using automatic dependent surveillance broadcast (ADS-B) data, with LSTM networks outperforming SVR in managing flow variations (Fig. 11). This aligns with broader trends where temporal modeling excels in dynamic ATM environments.

Figure 11. Architecture of the LSTM-based air traffic model [Reference Gui, Zhou, Wang, Liu and Sun48].

2.3 Quality control in aerospace manufacturing

In the rapidly evolving domain of aerospace, significant strides in ML, deep learning (DL) and computer vision (CV) are transforming aircraft inspection methodologies. This section explores the technical intricacies of algorithms deployed in quality control and visual inspection processes. High-resolution and sensitive charge-coupled device (CCD) cameras, combined with sophisticated imaging systems, serve as the foundational technology for this application [Reference Kumar and Kannan49]. A notable research initiative has developed an integrated system for automated aircraft surface inspection, utilising CCD cameras and CV algorithms. This system incorporates drones, tablets, and pan-tilt-zoom (PTZ) cameras to methodically survey surfaces, emphasising detailed defect identification – from hairline fractures to fluid leaks – rather than mere image acquisition [Reference Rice, Li, Ying, Wan, Lim, Feng, Ng, Jin-Li and Babu50].

Expanding beyond 2D imaging, envision a robotic inspector equipped with dual 2D cameras and a 3D scanner, designed to confirm assemblies match their 3D computer-adided design (CAD) models [Reference Ben Abdallah, Jovančević, Orteu and Brèthes51]. While routine inspections verify component presence and alignment, complexities like flexible part spacing or interference detection require 3D datasets. The employed CV algorithm involves three steps: handling parasitic edges, weighting edges and determining gradient orientations.

Further exploration reveals the use of ANNs to enhance continued operational safety (COS) for aircraft, spacecraft, and unmanned aircraft systems (UAS) [Reference Ghimire and Raji7]. Leveraging databases like service difficulty reports (SDR) and NASA, these ANNs predict certification-critical parameters via supervised learning, improving aerospace component reliability.

Beyond component inspection, spatial awareness systems rely on visual simultaneous localisation and mapping (VSLAM) and visual odometry (VO) to ascertain vehicle orientation and position [Reference Al-Kaff, Martin, Garcia, de la Escalera and Armingol52]. Concurrently, CV-driven defect categorisation – such as corrosion, cracks and punctures – streamlines maintenance workflows [Reference Das, Hollander and Suliman53].

At the forefront of CV advancements, convolutional neural networks (CNNs) like U-Net excel in defect detection. U-Net’s architecture combines contracting paths (context capture) with expanding paths (precise localisation), integrating convolutional and pooling layers for feature extraction and upscaling. This preserves spatial information, enabling effective segmentation in biomedical and aerospace visual analysis.

2.4 Autonomous navigation and control

Many DL frameworks designed for visual scene interpretation, navigation, guidance and control in UAS primarily utilise CNNs. Traditional AI problem-solving algorithms, often optimisation-based, are rooted in mathematical methodologies and can find near-optimal solutions for nondeterministic polynomial-time (NP)-hard problems.

Learning-based methods encompass both model-based AI algorithms and modern techniques, including reinforcement learning (RL) and deep reinforcement learning (DRL), which use Markov decision processes (MDP) or partially observable MDP (POMDP), as well as asynchronous advantage actor-critic (A3C) architectures for decentralised training.

While the AI sector expands rapidly, many effective methods for autonomous UAV navigation remain underexplored. Autonomous UAV navigation enhances flexibility in dynamic environments, relying on optimisation-based approaches such as particle swarm optimisation (PSO), ant colony optimisation (ACO), genetic algorithm (GA), simulated annealing (SA), pigeon-inspired optimisation (PIO), cuckoo search (CS), A^* algorithm, differential evolution (DE) and grey wolf optimiser (GWO). Researchers have adapted these algorithms for mission-specific constraints to achieve optimal results [Reference Rezwan and Choi54].

This survey bridges technical foundations with practical applications by outlining UAV characteristics and types to familiarise readers with architectures, summarising navigation systems and real-world implementations, and detailing the principles of AI algorithms used for autonomous navigation.

2.5 Image recognition and computer vision

Recent advancements in image recognition and CV have significantly impacted numerous industries, including aerospace. Leveraging sophisticated algorithms and ML techniques, these technologies analyse visual data to extract meaningful information from images and videos. In aerospace, CV is crucial for two primary applications: AI-driven remote sensing and quality inspection of components.

In remote sensing, AI enhances data processing efficiency and accuracy, automating object recognition/classification and detecting environmental changes for improved decision-making. As noted by Lary [Reference Lary55], AI’s ability to process geospatial data is instrumental here. Among AI’s six primary directions identified by Wu [Reference Wu, Xie, Lu and Zhu56] – CV, natural language processing, robotics, cognitive computing, game theory/ethics and ML – the latter is particularly vital. Wang et al. [Reference Wang, Yan, Mu and Huang57] emphasise ML’s necessity in advancing remote sensing capabilities.

For quality inspection, CV systems use cameras and algorithms to perform tasks traditionally reliant on human vision. The aerospace industry increasingly adopts intelligent visual inspection systems to ensure component quality. Advances in ML, especially DL, have enabled automatic solutions for inspecting internal/external components, from surface defects to structural integrity.

2.6 Maintenance and documentation

AI has become a pivotal force in aviation maintenance, offering innovative solutions across diverse applications. Inspection technologies leverage CNNs with autonomous drones to automate visual inspections, enhancing defect detection accuracy for issues like dents through image augmentations and pre-classification techniques [Reference Kabashkin, Misnevs and Zervina58]. Complementing this, ML and Internet of Things (IoT)-based methods predict thermal performance in aircraft wing anti-icing systems, outperforming traditional computational fluid dynamics with ANNs [Reference Abdelghany, Farghaly, Almalki, Sarhan and Essa59].

Beyond component-specific applications, AI drives sustainability efforts: environmental impact analysis integrates statistical and ML methods to assess fuel burn, emissions and noise, supporting data-driven sustainable aviation [Reference Gao and Mavris60]. For maintenance optimisation, DNNs detect corrosion in aircraft lap joints with human-level precision, enabling condition-based maintenance [Reference Doğru, Bouarfa, Arizar and Aydoğan39]. In rotorcraft, fuzzy neural networks and individual blade control (IBC) reduce hub vibrations, informing better control laws [Reference Yang, Gao, Wang and Ni61].

Predictive capabilities extend to core systems: generalised regression neural networks (GRNN) accurately forecast exhaust gas temperature baselines for engine health [Reference Wang and Zhao62]. ML models also advance lithium-ion battery research for aviation, focusing on health estimation and fault diagnosis [Reference Chen, Qi and Wang63].

Global aviation authorities recognise AI’s transformative potential. International bodies such as the International Civil Aviation Organisation (ICAO) emphasise AI training for professionals, while the European Union Aviation Safety Agency (EASA) proposes AI oversight frameworks and EUROCONTROL’s FLY AI report prioritises air traffic management adoption [64–66]. Regional initiatives, including the European Commission’s AI strategy, promote human-centric AI in transportation, while the Federal Aviation Administration (FAA) and International Air Transport Association (IATA) drive research and development and workforce readiness [67–69].

4.1 Space rovers

ML can be effectively utilised in the aerospace sectors to enhance mathematical computations. Rovers shown in Fig. 15, which are devices engineered for exploring the surfaces of planets and other celestial bodies, stand to benefit significantly from ML applications. In the context of rovers, ML algorithms facilitate a range of tasks including autonomous navigation, path planning and anomaly detection. Additionally, these algorithms are instrumental in mechanical applications such as structural analysis, materials selection, design optimisation, fault detection and diagnostics, among others. A notable study explored the implementation of a ML technique known as Q-Learning, detailed in Ref. (Reference Gankidi and Thangavelautham77). Q-Learning is a model-free reinforcement learning algorithm where an agent learns optimal policies by iteratively updating a Q-table that maps state-action pairs to expected cumulative rewards. The rover interacts with its environment, receiving rewards for desirable behaviours (e.g., avoiding obstacles, conserving energy) and penalties for undesirable ones. Through trial-and-error interactions, it maximises the long-term reward signal defined by Equation (1), as illustrated for the rover in Fig. 15:

(1) \begin{align}Q\left( {{s_t},{a_t}} \right) \leftarrow Q\left( {{s_t},{a_t}} \right) + \alpha \left[ {{r_{t + 1}} + \gamma \mathop {{\rm{max}}}\limits_a Q\left( {{s_{t + 1}},a} \right) - Q\left( {{s_t},{a_t}} \right)} \right]\end{align}

where $\alpha $ is the learning rate, $\gamma $ the discount factor, and ${s_t}$ , ${a_t}$ , ${r_{t + 1}}$ represent the state, action and reward at time $t$ . This approach is particularly suited for rover navigation in unpredictable extraterrestrial environments, where predefined rules may fail. The field-programmable gate array (FPGA) implementation enables hardware acceleration of these computations, critical for real-time decision-making under resource constraints. Specifically, Q-learning has been applied in rovers to improve decision-making and navigation strategies. The research also demonstrated the deployment of Q-learning in both single neuron and multilayer perceptron architectures on a Xilinx Virtex 7 FPGA.

Figure 15. Mars rover’s robotic arm drove a drill bit into flat patch of the rock [Reference Bajracharya, Maimone and Helmick78].

Concurrently, another study focused on the locomotion challenges of tensegrity robots, which are utilised for planetary exploration, as described in Ref. (Reference Zhang, Geng, Bruce, Caluwaerts, Vespignani, SunSpiral, Abbeel and Levine79). This research proposed an advanced version of the mirror descent guided policy search (MDGPS) algorithm to autonomously learn effective locomotion gaits, addressing the intricate dynamics of these robots. Furthermore, an innovative approach to identifying rare geological phases on Mars was investigated, employing a Bayesian classifier trained with spectral data from the Compact Reconnaissance Imaging Spectrometer for Mars (CRISM), as outlined in Ref. (Reference Dundar, Ehlmann and Leask80). This classifier aids in the detection of unique geological features, showcasing the broad applicability of ML in space exploration contexts. Building on this, further explorations into the Jezero crater have revealed jarosite and silica on its floor, while chlorite-smectite and Al-rich phyllosilicates have been detected along the crater walls. These mineralogical findings suggest a complex history of water activity within the Jezero region. Moreover, the application of ML techniques has proven to be significantly more effective than traditional methods for such reconnaissance, underscoring the potential of ML to enhance geological analysis in planetary exploration.

In a study conducted by Neil et al. [Reference Abcouwer, Daftry, Del Sesto, Toupet, Ono, Venkatraman, Lanka, Song and Yue81], a novel pathfinding algorithm for navigation across uneven terrains was developed. The algorithm employs a heuristic known as the gradient convolution, which integrates a metric called the angular cost estimate (ACE). ACE quantifies the irregularity of the terrain beneath the rover’s wheels. The approach utilises a deep convolutional neural network (DCNN) architecture akin to the UNet model. This architecture processes height maps to produce ACE cost maps, thereby optimising computational efficiency. The implementation achieved an overall accuracy of 95.3%. This example illustrates the application of ML techniques to enhance performance and reduce computational demands in scenarios where traditional models are inadequate.

4.2 Satellites and earth observation

In the expansive realm of outer space, satellites are integral to global connectivity. The incorporation of AI and ML technologies is revolutionising the utilisation of satellite systems as shown in Fig. 16. This section explores the application of ML and DL across various satellite operations. Geostationary Earth Orbit (GEO) satellites, predominantly utilised for communication services such as broadcasting, broadband Internet access and telecommunications, benefit significantly from these advancements. Research in this domain has employed models including SVMs, DNNs and DQNs to optimise resource allocation. Such innovations not only foster more efficient system designs but also contribute to substantial cost reductions [Reference Ortiz-Gomez, Lei, Lagunas, Martinez, Tarchi, Querol, Salas-Natera and Chatzinotas82].

Figure 16. AI chip implementation in GEO satellite [Reference Ortiz-Gomez, Lei, Lagunas, Martinez, Tarchi, Querol, Salas-Natera and Chatzinotas82].

In low Earth orbit (LEO) constellations, which are utilised for communication and Earth observation with hundreds to thousands of satellites, monitoring each satellite’s state poses challenges. Autoencoders extract features from parameters like position, velocity and mission type, while ML classifiers (random forest (RF), SVM, neural networks) identify orbit states for debris mitigation and spectrum allocation [Reference Wang and Li83]. Remote sensing applications further demonstrate ML’s versatility: MIDAPS-AI combines decision trees, neural networks and SVM on LEO/GEO satellites for disaster management and debris detection [Reference Maddy and Boukabara84], while nanosatellites like OPSAT implement online ML for fault detection, isolation, and recovery (FDIR), enabling real-time analysis in resource-constrained environments (Fig. 17) [Reference Labrèche, Evans, Marszk, Mladenov, Shiradhonkar, Soto and Zelenevskiy85].

Figure 17. Orbit-AI app implementation in OPSAT [Reference Labrèche, Evans, Marszk, Mladenov, Shiradhonkar, Soto and Zelenevskiy85].

High-throughput and navigation systems also leverage ML: very high throughput satellites (VHTS) optimise power/bandwidth using elevation and traffic data [Reference Ortiz-Gómez, Tarchi, Martinez, Vanelli-Coralli, Salas-Natera and Landeros-Ayala86], while GNSS apply ML for signal processing and anomaly detection [Reference Siemuri, Kuusniemi, Elmusrati, Välisuo and Shamsuzzoha87]. Cross-domain challenges – resource allocation, interference, latency – are addressed via neural networks, SVM, and genetic algorithms, enhancing Quality of Service (QoS) [Reference Homssi, Dakic, Wang, Alpcan, Allen, Kandeepan, Al-Hourani and Saad88, Reference Henarejos, Vázquez and Pérez-Neira89].

Data privacy and satellite health are critical considerations: Federated Learning (FedSpace) enables GDPR-compliant distributed training (Fig. 18) [Reference So, Hsieh, Arzani, Noghabi, Avestimehr and Chandra90], while CNNs analyse temperature sensors for real-time health diagnostics, outperforming traditional methods in noise resilience [Reference Saudi, Hussein, Said and Al-Zaidy91].

Figure 18. Overview of aggregation schedular in FedSpace [Reference So, Hsieh, Arzani, Noghabi, Avestimehr and Chandra90].

Satellite data analytics rely on ML for tasks like classification and imagery analysis: SVM, decision trees and random forests handle regression tasks, while CNNs, RNNs and LSTM networks process imagery and fusion (Fig. 19) [Reference Sisodiya, Dube and Thakkar92]. Future advancements in machine learning operations (MLOps) and space-grade processors will enable adaptive AI for dynamic space environments [Reference Veyette, Aylor, Stafford, Herrera, Jumani and Lineberry93].

Figure 19. Application of data collected by satellites to monitor construction rate using segmentation [Reference Sisodiya, Dube and Thakkar92].

4.3 Deep space, autonomous health monitoring in space

The mental and physical demands of multi-year space missions necessitate enhanced astronaut medical care, and AI shows great promise in improving future crew support systems. By integrating data from sensors monitoring heart rate, skin temperature, exercise and sleep patterns, AI-powered predictive health analytics can provide customised interventions tailored to each astronaut. This holistic approach, combining real-time vital signs, behavioural indicators and environmental conditions, enables sophisticated diagnostics, early risk warnings and personalised treatment plans. For instance, the Crew Interactive Mobile Companion (CIMON) as shown in Fig. 20, an AI robot designed by Airbus, IBM and the German Aerospace Centre, assists astronauts on the International Space Station (ISS). CIMON can navigate the station, document experiments and offer procedural guidance, while also providing emotional support by sensing stress levels and guiding astronauts through therapeutic exercises. Future intelligent systems on the ISS and lunar Gateway will further anticipate astronauts’ needs and automate tasks, with AI virtual assistants adapted for psychological support during Mars missions, where communication delays with ground control are significant [Reference Loeffler94].

Figure 20. CIMON with ESA astronaut Alexander Gerst (Credit: ESA).

Scientific research aboard the ISS is critical for advancing space medicine: NASA’s SpaceX Crew-6 recently launched and docked at the ISS, initiating over 200 experiments. These include the Cardinal Heart 2.0 project, which examines whether approved drugs mitigate microgravity-induced changes in heart-cell function, and microbial studies sampling ISS life-support vents to analyse microorganism behaviour in space. Complementing this, NASA’s Biosentinel experiment – deployed on a CubeSat in heliocentric deep space radiation effects, marking the first such study in five decades. To expand research for missions beyond LEO, including lunar and Martian missions, highly automated approaches with AI and ML will be pivotal. NASA’s 2021 workshop explored AI’s role in advancing space biology and personalised healthcare [95].

AI has significantly transformed expert medical systems, particularly in differential diagnoses. DL advancements enable AI systems to match human experts in diagnosing diseases across radiology and pulmonology. Deep CNNs excel in extracting intrinsic features from raw data, achieving superior results in medical image tasks like classification and segmentation. For instance, OpticNet-71 trained on optical coherence tomography (OCT) images identifies age-related macular degeneration, diabetic macular edema, drusen and choroidal neovascularisation, while CheXNeXt diagnoses 14 diseases from chest radiographs. Semantic segmentation utilises architectures like U-Net for retinal vessel and brain-tumor segmentation. Generative adversarial networks (GANs) further innovate medical imaging: conditional GANs generate fluorescein angiography images from Color Fundus, and RV-GAN specialises in retinal vessel segmentation. These advancements underscore AI’s potential to revolutionise clinical support systems [Reference Waisberg, Ong, Paladugu, Kamran, Zaman, Lee and Tavakkoli96].

AI in space medicine faces unique challenges: sparse astronaut data for training, limited prospective research in space environments and ethical/legal complexities like managing astronaut-doctor relationships and informed consent. These necessitate specialised approaches and ethical frameworks for effective integration [Reference Waisberg, Ong, Paladugu, Kamran, Zaman, Lee and Tavakkoli96].

4.4 Launch vehicles

Recent advancements in AI and ML have significantly enhanced rocket technology reliability and performance. Solid rocket motors (SRMs), critical for launch vehicles and defense missiles, benefit from AI-driven defect detection. Liu et al. [Reference Liu, Sun and Miller97] pioneered sensor systems using CNNs and LSTM networks to identify inner bore cracks and propellant delamination during testing. Building on defect detection, ML algorithms address propellant combustion unpredictability: Debus et al. [Reference Debus, Ruettgers, Petrarolo, Kobald and Siggel98] analysed fuel-oxidiser interactions with RF, SVM, K-Nearest Neighbour (KNN) and ANNs, demonstrating RF’s superior accuracy over traditional methods.

Hybrid rocket propellant analysis further exemplifies AI’s versatility: Surina et al. [Reference Surina, Georgalis, Aphale, Patra and DesJardin99] combined imaging techniques (thresholding, last image subtraction, spatial filtering) with U-Net architectures and Monte Carlo dropout to measure regression rates, enhancing performance insights. For flight performance evaluation, surrogate neural networks (SNNs) simulate hybrid rocket engine (HRE) dynamics, heat transfer and thrust parameters with precision comparable to full-scale simulations [Reference Zavoli, Zolla, Federici, Migliorino and Bianchi100].

Real-time telemetry optimisation leverages CNNs: Li et al. [Reference Li, Peng, Zhao, Yang and Wang101] streamlined rocket status assessment using CNN architectures (input, convolution, pooling, fully connected, output layers) for video data processing. Collectively, these advancements – defect detection, combustion analysis, regression measurement, performance simulation and telemetry – enhance safety, reliability and efficiency in rocketry.

4.5 AI’s role in Chandrayaan-3 mission

AI played a pivotal role in enhancing navigation, hazard avoidance, data interpretation and enabling autonomous decision-making in the Chandrayaan-3 mission. Chandrayaan-3 (Fig. 21), the successor to Chandrayaan-2, successfully touched down on the Moon’s southern pole on August 23, 2023 [102]. Building on discoveries from Chandrayaan-1 (water detection in the Aitken basin) and Chandrayaan-2 (mineral mapping via orbiter data), AI systems managed critical landing phases by integrating altimeters, velocimeters and cameras to adjust altitude, fire thrusters and scan for obstacles [103–Reference Verma105].

Figure 21. Chandrayaan-3 (Credit: ISRO).

Central to the landing was the terrain relative navigation (TRN) system: using a camera and onboard computer, TRN matched lunar surface images with pre-loaded maps, analysing terrain features, slopes and hazards via image processing and pattern recognition. While exact models are undisclosed, CNNs, Siamese networks or LSTM networks likely enabled feature comparison and sequential data processing during descent [Reference Isaac106].

The navigation, guidance, and control (NGC) system dynamically adjusted speed, orientation, and trajectory using TRN inputs and mission objectives. Potential control mechanisms included proportional-integral-derivative (PID) controllers, DQNs trained on Chandrayaan-2 data, DNNs for real-time sensor adjustments or RNNs for sequential data handling.

For the Pragyan rover (Fig. 22), AI enabled autonomous navigation: Likely strategies included RL, A^*/rapidly exploring random trees (RRT) for path planning, and CNNs for real-time image analysis. Transfer learning may have adapted pre-trained models to lunar conditions. Instruments like the alpha-proton X-ray spectrometer (APXS) and laser-induced breakdown spectroscope (LIBS) utilised AI for geological analysis – CNNs identified features, regression/classification interpreted spectral data and matching algorithms cross-referenced mineral libraries [Reference Kashettar107].

Figure 22. Pragyan – lunar rover (Credit: ISRO).

4.6 Current work by space agencies

Transitioning from exploratory research to practical space applications may seem daunting, but the European Space Agency (ESA) is already utilising AI in its missions. Rovers, for instance, can autonomously navigate obstacles, and data downloads from Mars rovers are now scheduled using AI. AI also supports astronauts aboard the ISS, enhancing their daily operations.

ESA’s Hera planetary defense mission exemplifies AI’s potential, autonomously navigating through space toward an asteroid by fusing sensor data and making real-time decisions, much like self-driving cars. While most deep-space missions rely on human controllers, Hera’s onboard autonomy sets a new standard. Satellites are also gaining more autonomy to perform collision avoidance manoeuvers amidst increasing space debris. In January 2021, ESA and the German Research Centre for Artificial Intelligence (DFKI) launched ESA_Lab@DFKI, focusing on AI systems for satellite autonomy and collision avoidance.

AI also enhances satellite navigation, improving weather forecasting and detecting rogue drones in sensitive airspace. ESA-led projects have applied AI for autonomous situational awareness in ships, and the German Aerospace Centre (DLR) has developed AI methods for space and Earth applications, establishing an Institute of Artificial Intelligence Security in 2021. DLR’s AI assistant, CIMON, supports astronauts on the ISS by performing voice-controlled tasks and navigating the station.

NASA also employs AI extensively, with an artificial intelligence group dedicated to supporting scientific analysis, spacecraft operations, mission analysis, deep space network operations and space transportation systems. NASA’s cognitive radio technology ptimises communication networks by using ‘white noise’ areas in communication bands, enhancing the efficiency of limited telecommunication resources. AI has also improved solar research data and facilitated autonomous spacecraft operations, reducing communication delays. NASA’s collaboration with Google led to the discovery of two exoplanets by training AI algorithms to analyse Kepler mission data. This success prompted the use of AI in NASA’s TESS mission to identify potential exoplanets.

Japan’s Japan Aerospace Exploration Agency (JAXA) pioneered AI integration in space missions with its Epsilon rocket, which autonomously performs performance checks, simplifying payload launches. JAXA’s ‘Int-ball’ robot autonomously captures ISS experiment photos, saving astronauts time. The French space agency CNES, in partnership with Clemessy, optimised rocket tank filling using AI neural networks. The UK Space Agency funded AI projects to detect buried archaeological remains via satellite imagery, and the Italian Space Agency co-founded an AI-focused company. These initiatives highlight the transformative potential of AI in space exploration and operations, underscoring the importance of ongoing research, development and international collaboration [108].