As space exploration missions target more distant planetary bodies in our solar system–including Ocean Worlds (e.g., Titan, Europa, Enceladus), comets (e.g., 67P/Churyumov-Gerasimenko), and asteroids (e.g., Bennu, Ryugu, Psyche) – it is becoming clear that onboard autonomy will play an increasingly critical role in optimizing the science achieved. The Planetary and Astrobiology Decadal Survey (National Aeronautics and Space Administration, 2022) states that “missions targeting remote planetary bodies will face limited data rates and volumes and will require onboard science autonomy to optimize science return.” Future missions to ever more distant targets will be severely limited by stringent communication constraints including narrow communication time windows, long latencies, and restrictive bandwidth, which renders prohibitive any ambitious missions that rely on current operation models based on pre-programmed data acquisition interleaved with ground-in-the-loop interactions. In addition, these future missions would be further constrained in their ability to respond optimally to uncertain and new environments and events under such a constrained cadence.

Groundbreaking planetary missions such as Dragonfly targeting Titan in the 2030s (Barnes et al., 2021; Turtle et al., 2020), Europa Clipper on its way to Europa (Phillips and Pappalardo, 2014), Rosalind Franklin destined for Mars (Vago et al., 2017) and the Enceladus Orbilander (identified as a priority in National Academies of Sciences Engineering and Medicine (2022), Mackenzie et al. (2021)), include advanced science instruments able to collect greater data volumes than can be sent back to Earth. These missions aim to conduct in situ compositional analysis on planetary bodies where data downlink is highly constrained. Mass spectrometers (MS) are powerful instruments onboard Dragonfly, Clipper, and Rosalind Franklin (and potentially included on Orbilander), that aim to identify unknown compounds via their mass-to-charge ratio and perform quantitative analysis of these compounds. MS instruments have been essential to space missions over the past several decades (e.g., Pioneer Venus, Viking, Galileo, Cassini, Mars Science Laboratory) with several more en route (e.g., JUpiter ICy moons Explorer (JUICE), Europa Clipper) or under development (e.g., Rosalind Franklin, Dragonfly) (Chou et al., 2021). Instruments currently being developed for future missions are capable of generating even more data (Arevalo et al., 2019) with sampling rates already increasing from 50 samples per second for the quadrupole mass spectrometer Sample Analysis at Mars (SAM) onboard Curiosity, to 50,000 samples per second for the MOMA ion trap mass spectrometer on ExoMars, and expected to reach millions per second for next-generation orbitrap mass spectrometers. This rapid growth far exceeds the data transmission rates to Earth which are limited by physics and communications technologies, emphasizing a widening gap between the volume of data that can acquired compared to what can be transmitted back to Earth. Therefore, major advancements essential to support future mission operations and data collection could be enabled with more autonomy.

Given how autonomy could enhance and enable space exploration missions, we investigate the concept of science autonomy–the ability of a science instrument to analyze its own data, make decisions about what data to send back to Earth, and learn to adapt itself to improve the next measurements–for a Mars exploration mission as an initial proof-of-concept. Machine learning (ML) is a promising tool to enable time-efficient data triage and classification required by onboard science autonomy (Chien and Morris, 2014; Thompson et al., 2015; Chien et al., 2017). In this context, autonomy depends on the ability to prioritize, filter, and interpret instrument data so that the most informative data are transmitted to Earth and used to guide subsequent operations. We leverage our previous work with the MOMA instrument on the Rosalind Franklin (ExoMars) mission (Da Poian et al., 2022b) using ML and data science tools to further investigate the application of science autonomy on mass spectrometry data for space missions. While full science autonomy would ultimately require algorithms onboard, our initial implementation focuses on supporting the science team’s decision-making process on Earth during operations on Mars by enhancing data interpretation–a step toward developing and validating approaches that can eventually transition to onboard use. Achieving this level of autonomy could highly benefit from robust ML algorithms as they offer a promising solution for scientific data analysis during constrained space operations. However, the development of reliable ML models in this context is limited by the scarcity of representative training datasets. Mass spectrometers on space missions are optimized for their specific target environments, making them unique and expensive to build and operate, with typically only two or three instances of the instrument built per mission. This limited availability restricts opportunities to collect diverse and statistically significant datasets for ML models development. To address this limitation, our work investigates the use of data augmentation techniques to generate scientifically consistent mass spectra to expand the limited available dataset volumes without introducing bias. By carefully, designing and validating these generated data products in collaboration with instrument experts, we aim to create training datasets that are both relevant for the purpose of ML model development and suitable to advance future onboard autonomy.

Data science is an interdisciplinary field that extracts insights and knowledge from data using various techniques such as statistical analysis and data mining (Cao, 2017). Data science uses mathematical and statistical theory (often in combination with programming) to describe data relationships and trends. ML is a subset of artificial intelligence (AI) that focuses on algorithm development to enable computers to learn and improve from experience (Figure 1). ML algorithms can mainly be divided into unsupervised ML techniques, aiming at finding unbiased patterns and similarities in the input dataset to group similar data into clusters (e.g., dimensionality reduction and clustering algorithms), and supervised ML techniques, aiming at using input labels to train the model to predict a specific output (e.g., classification and regression algorithms) (Carbonell et al., 1983; Ayodele, 2010).

ML and data science have become widely investigated and adopted across both the science community and the space sector. These methods are tackling many challenges, from the limited communication bandwidth and delayed data transmission, to the ever-expanding volumes of scientific and engineering data. For instance, ML algorithms are used for autonomous navigation and hazard avoidance on planetary surfaces (e.g., Mars rovers (González et al., 2020), real-time anomaly detection onboard spacecraft (Hundman et al., 2018), onboard science event detection (Wagstaff et al., 2019), and mission planning optimization (Wang et al., 2022). Deep learning approaches have opened new frontiers in space exploration such as in the classification of Martian terrains (Rothrock et al., 2016) as well as in the automatic detection of exoplanets from Kepler and TESS light curves (Malik et al., 2022). Collectively, these applications highlight the growing importance and critical role of ML algorithms in enabling more autonomous space exploration. Building on this trend, a natural step is to extend ML-driven autonomy to science instruments themselves. In the case of MS, ML methods could enable onboard science autonomy by triaging large datasets, prioritizing scientifically valuable features for downlink, and adapting measurement strategies in near-real time to maximize mission science return. As the development of such onboard capabilities through ML algorithms is constrained by the scarcity of representative training data from flight-like instruments, our work investigates data augmentation approaches to generate realistic, scientifically valid MS data to expand limited datasets without introducing bias.

In this work, we focus on advancing ML applications for mass spectrometry data from Mars by leveraging datasets from the engineering test unit (ETU) of the Mars Organic Molecule Analyzer (MOMA) instrument planned to be onboard the Rosalind Franklin (ExoMars) rover mission (Goesmann et al., 2017; Brinckerhoff et al., 2013) (Figure 2). MOMA is a dual-ionization source mass spectrometer instrument designed to detect and characterize a broad range of organic molecules in surface and subsurface samples. MOMA is centered around a linear ion trap mass spectrometer (ITMS) interfaced with two complementary “front-end” analytical techniques: laser desorption/ionization (LDI) and gas chromatography (GC). This dual capability enables MOMA to target a broad range of organic molecules (with masses up to 1,000 Da), from volatile to nonvolatile compounds. Here we only focus on laser desorption/ionization mass spectrometry (LDMS) mode data as it is particularly valuable for the detection of nonvolatile complex organics over a wide range of molecular weights, providing insights into potential biosignatures or prebiotic chemistry on Mars (Li et al., 2017). By leveraging this LDMS dataset, we aim to advance analytical techniques to maximize scientific return.

Our work aims to mature the proof-of-concept science autonomy approaches we previously developed (Da Poian et al., 2022a; Do et al., 2024) and refine ML techniques on this LDMS dataset to enhance MOMA’s scientific return. Our ML task focuses on clustering similar spectra and on the classification of mass spectra into main chemical families (10 different families/labels defined by MOMA scientists) and predicting chemical sample composition (67 labels) of new, previously unmeasured mass spectra during Mars exploration. This work provides a data analysis pipeline and internal tools to allow the MOMA science team to efficiently assess the chemical composition of Martian samples. The pipeline aims to enable scientists to quickly analyze new mass spectra, guiding operational decisions such as performing tandem mass spectrometry (i.e., MS/MS), isolating peaks of interest, adjusting laser desorption/ionization parameters, and determining if a gas chromatography mass spectrometry (GCMS) follow-up experiment is worthwhile, ultimately accelerating scientific processing in time-limited mission operations.

By using the MOMA instrument as a proof-of-concept, we are demonstrating initial capabilities that, while currently developed for Earth-based spectral analysis and interpretation, represent critical steps toward enabling science autonomy on future planetary missions. These methods could be generalized and applied to other planetary mission concepts to enable different levels of autonomy based on the mission needs and instruments capabilities (Azari et al., 2021). A major challenge in developing such ML-based autonomy lies in the limited availability of representative training data from flight-like instruments. Here, we specifically investigate techniques to increase the amount of data available for ML algorithms training and evaluate how dataset volumes impact ML algorithms performance. At this stage, all ML model training in our research is performed on Earth, where computational resources enable to systematically explore how data augmentation affects learning. However, our long-term vision is to extend these approaches toward in situ adaptability–where models deployed onboard a spacecraft could be retrained or fine-tuned using newly acquired data. Such capabilities would enable context-aware decision-making and instrument optimization during resource-limited missions operations. In this context, exploring shorter training durations (fewer epochs) serves not as an operational constraint but mimics constrained learning conditions due to flight-like conditions, allowing us to assess how augmentation enhances model generalization when data or training iterations are limited. By systematically evaluating how these generated datasets affect ML performance, we lay the groundwork for more robust and scalable ML solutions that can ultimately support autonomous MS operations in future planetary science missions.

Our work leverages mass spectral data from the ETU of the MOMA instrument (Figure 2) (Li et al., 2017). The ETU is a form, fit, and functional duplicate of the flight model (FM) that permits analysis of a wide range of standard and Mars analog samples into a proxy carousel for direct LDMS analysis using a UV (266 nm) pulsed laser. The analysis performed on the ETU represents a high-fidelity analog of MOMA operation and data production. This instrument is a critical flight resource for the Rosalind Franklin mission and is thus tracked closely and maintained in a controlled configuration and high level of cleanliness, with significant operator support requirements; as such, the range and scope of new samples analyzed by the ETU is limited. To date, several hundred samples have been analyzed on the ETU, each producing thousands of individual mass spectra on average.

Our original dataset is composed of about ∼30,000 LDMS spectra collected on the ETU by the science team for scientific purposes (Da Poian et al., 2022b), with engineering test data excluded and archived separately in an internal NASA database. The data augmentation we investigate here involves processes to artificially increase the available dataset by introducing statistically plausible variations to the original data. While data augmentation techniques are commonly applied to image and signal processing domains (Shorten and Khoshgoftaar, 2019; Wang and Perez, 2017), augmentation for MS data, especially in planetary science context, requires careful consideration to preserve the underlying scientific validity of the data. Several challenges arise when designing augmentation strategies for MS data such as: -

The preservation of scientific validity: as each peak represents a specific ion or fragment, the alteration of a peak intensity or position could lead to the wrong interpretation of the chemical signature.

To mitigate these challenges, we developed traditional augmentation techniques that carefully constrain modifications of MS data within chemically meaningful ranges, ensuring that the generated spectra remain representative of the true instrument behavior. For instance, intensity changes were applied globally rather that to specific isolated peaks, shifts were bounded within instrument resolution limits, and noise additions were derived from experimental runs. Our augmentation strategies maintain fidelity to the scientific constraints of LDMS data while still producing a more diverse and robust dataset for input to ML algorithms.

We developed scripts for various data augmentation techniques (Table 1) designed to mimic realistic mass spectral variability due to instrumental or environmental factors. These scripts include techniques to shift peaks by shifting m/z values (i.e., the x-axis values) and the relative intensity values (i.e., the y-axis values), as well as adding noise (focusing on random noise). We also investigate a stretch transformation and the shift of some peaks. Each technique is assessed by MOMA scientists for its consistency with the physical principles underlying LDMS data and more specifically MOMA-related chemistry. The goal of scientific validity is to ensure that the transformations do not introduce artifacts inconsistent with MS science. This augmentation strategy aims to enhance the ML model generalization and robustness while preserving the scientific validity of the augmented mass spectra. We introduce below the various augmentation techniques we implement in a Python script on the original MOMA dataset.

Once the augmented MS data were created, we implemented a validation process in close collaboration with the MOMA science team to ensure their scientific consistency related to the MOMA instrument. The MOMA science team reviewed and analyzed the augmented spectra using similar diagnostic process as for flight data analysis. They specifically focused on: -

the peak shapes, consistent with the MOMA instrument,

These criteria were assessed through manual expert review process using visualization tools applied to MOMA ETU data. This process was iterative to refine data augmentation characteristics to ensure that the augmented datasets maintained the science integrity and could be used to support the ML model development without introducing bias, artifacts or misleading features.

In this work, we build upon earlier efforts that developed ML-based analyses of MOMA LDMS mass spectra (Da Poian et al., 2021). The original dataset consists of LDMS spectra collected from the MOMA ETU model during laboratory experiments using both individual chemical standards and complex mixtures representative of analog materials relevant to Mars exploration. These laboratory experiments covered a range of molecular families expected during MOMA flight operations, providing a diverse and representative training set for ML model development (Da Poian et al., 2022a). In collaboration with the MOMA science team, each mass spectrum was labeled using two different ways: a category label, representing the chemical family of the tested compound (e.g., “pure mineral”, “pure organic standard”), and a sample label, representing the specific physical sample analyzed (e.g., “hematite”). The dataset is moderately imbalanced with calibration spectra being the most frequent class due to repeated tuning and performance checks on the ETU model. Each mass spectrum is represented as a one-dimensional intensity array, with both associated experimental metadata and labels compiled in a unified CSV file referred to here as the “original dataset”.

The various data augmentation methods were applied to this original dataset in custom (Table 1) generating new mass spectra based on the original data. We then combine these generated spectra with the original data to form multiple training configurations with various percentages (10%, 30%, 50%, 70%, and 90%) of augmented data. In addition to varying the amount of augmented data, we also vary the proportion of calibration data (cesium iodide) used in the input data – 0%, 50%, or 100%. These various input datasets create multiple training configurations. Finally, we varied the number of training epochs–the number of times the entire training dataset is passed through the model during learning–using 5, 10, 50, 100, and 200 epochs across all configurations of input data. The various options for the combinations of input data (i.e., amount of calibration data and amount of augmented data) and the various epochs values are represented in Figure 4.

To evaluate the impact of augmented data on the model performance, we leverage the NN framework previously developed in Da Poian et al. (2022b) that uses open-source Python libraries including Scikit-Learn, TensorFlow, and Keras, on the various combinations of input datasets (Figure 5). In brief, this NN uses a fully connected feed-forward architecture designed to classify mass spectra into predefined chemical categories called “category labels”. Each spectrum is represented as a unidimensional array by indexing data to rounded, “integerized” (transformed into integers corresponding to the nominal unit mass) values ranging from 1 to 2,000. In this work focusing on the investigation of data augmentation on ML model performance, this previously-developed NN is trained independently on each dataset combination (e.g., original, augmented) to predict the category label. By comparing model performance across these configurations, we quantify how augmentation strategies influence the NN’s ability to generalize to unseen data and to robustly identify scientifically meaningful features in mass spectra, using standard classification metrics.

When developing a ML classification model–such as predicting the chemical composition of samples from planetary bodies–it is essential to select and use appropriate performance metrics to evaluate the quality of the model’s predictions that reflect both statistical accuracy and scientific relevance. In the context of planetary MS data, these metrics help evaluate how reliably the model can distinguish between different chemical compound labels. In our case, a false positive, for instance, could represent a spectrum incorrectly classified as organic when it is not–potentially leading to an unnecessary activation of follow-up experiments (e.g., GCMS experiment) or in the future an incorrect prioritization of data for downlink. On the other side, a false negative could correspond to a missed detection of a scientifically significant compound, leading to lost science opportunities. Understanding these errors is critical for future autonomous decision-making frameworks. We summarize the most commonly used classification performance metrics below: -

Accuracy: the ratio of correctly predicted instances to the total instances in the dataset. While accuracy is a straightforward and widely reported metric, it could be misleading in the presence of class imbalance–a situation where some categories are represented by many more samples than others–or for specific tasks where the detections of false positives and false negatives carry different consequences.

As our dataset is moderately imbalanced–with some chemical classes (e.g., calibration sample) represented by many more spectra than others–single-number metrics (such as overall accuracy) can be misleading. To provide a meaningful assessment of the overall performance, the choice of the metrics needs to account for both class imbalance and scientific priorities.

The choice of evaluation metrics to use depends on the specific task to accomplish, the available dataset (e.g., class distribution), and the scientific or operational priorities of the task (e.g., relative importance of false positives and false negatives in the problem’s context). In our case, we are particularly interested in minimizing the number of false positives, as incorrectly predicting the presence of certain chemical compounds could lead to misleading scientific interpretations or mission operations decisions. At the same time, detecting rare but scientifically significant compounds–such as potential biosignatures or key chemical markers–is critical. For this reason, we use macro recall to evaluate the model’s ability to detect all classes equally, including low-frequency compounds, and macro f1-score to provide a balanced measure of precision and recall across all classes. Additionally, we carefully look at the weighted f1-score to capture overall model performance despite the presence of class imbalance while still reflecting the trade-off between false positives and false negatives in a way that aligns with planetary science and astrobiology science goals. All the plots for these three metrics for all the configurations are summarized in Figure 6.

Obviously, epochs are the dominant driver of computational cost represented by the training time (Figure 7). Augmenting the training rounds from 5 epochs to 200 epochs, increases the training time by about ∼102 min. Moreover, augmenting the size of the input dataset (through the augmentation % variable) substantially increases runtime as well. A jump from the original dataset (0% augmentation) to 90% augmentation increases the training time by about ∼60 min. Finally, including CsI in the input dataset has negligible impact on the training time.

As augmentation increases both the NN performance and its training time (i.e., computational cost), we can optimize it for Pareto efficiency described in the following section.

The Pareto frontier (or Pareto front) is used in multi-objective optimization and aims at describing the set of solutions that are optimal and non-dominated, i.e., improving one objective results in a loss in another (Navon et al., 2021). In our case, the two objectives are to maximize the NN performance (i.e., highest macro recall, macro f1-score, weighted f1-score) while minimizing the computational cost (i.e., training time). As shown in Figure 8, some of our configurations achieve excellent performance, but require long training times. The Pareto front identifies the configurations that offer the best trade-offs between speed and performance. Not only are specific configurations more efficient in terms of performance vs. training time, but theyalso provide insight into the generalization and robustness of the trained NN. In ML problems, generalization is the ability of a model to maintain high performance on unseen data, and robustness is the ability to remain stable across different data distributions or noise conditions.

Figure 8 shows that runs with few epochs and no augmentation complete rapidly but reach lower performance. Interestingly, for the 0% CsI experiment, the model trained with 50 epochs and no augmentation (denoted as “50e, 0%” on Figure 8) lies on the Pareto front and achieves near-optimal performance across the three metrics. For the 50% and 100% CsI experiments, the optimal Pareto solutions are for models trained with moderate augmentation (50%–70%) and 50 to 100 epochs. This suggests that a combination of high-quality calibration and controlled augmentation yields to a more efficient model. Finally, more extreme configurations (e.g., 200 epochs with 90% augmentation) reach marginal performance gains while requiring a large computational cost (e.g., 3–5 h).

Table 2 summarizes all the effects from varying a single variable (augmentation % in input dataset, CsI %, or number of epochs), from the interactions between two of these three variables, and effects of changes with the three variables.

In our study, high augmentation improves the generalization by exposing the NN to a broader variety of training examples, reducing the risk of overfit to the original dataset. In contrast, high CsI percentage (high-quality calibration data) helps the model’s robustness without overwhelming the computational cost. More epochs allow the model to refine the decision boundaries for the feature–label relationships, but the model tends to overfit after ∼100 epochs.

The Pareto front highlights the trade-off between classification performance and computational cost. In our configurations, the Pareto frontier between performance and training cost tends to be for configurations with 50%–70% augmentation, 50–100 epochs, and ≥50% CsI.

Our study aimed at investigating data augmentation techniques for MS data with the goal of preserving scientific interpretability to help in NN learning for planetary science applications. This work directly addresses one of the main barriers to enable ML-driven science autonomy–the scarcity of representative training data from flight-like instruments. We first investigated various ways to augment our original dataset, and we established their scientific validity through a verification process conducted in close collaboration with instrument domain experts. We then assessed how these augmented spectra, calibration data, and training duration affect NN performance for our specific MS classification task.

Across all metrics, the use of augmented data consistently improved generalization, but the magnitude of improvement depended on the training regime. The strongest gains happened for lower epochs where augmentation effectively compensates for the limited training time. For intermediate epochs values (50 epochs), the improvement was moderate, and performance gains plateaued at higher epochs (100 epochs and above). Augmentation most effectively boosts NN learning under specific conditions, namely, when computational resources and training time are limited.

Similarly, the amount of calibration data used in the training set also had a measurable effect on the NN performance: including some (50%) or all (100%) of the CsI generally improved all metrics, especially for higher augmentation levels. Finally, training for more epochs increased performance with the most noticeable improvements from 5 epochs to 50 epochs. Our Pareto front analysis shows that beyond ∼100 epochs, the training time increases while the performance returns are plateauing, especially for high augmentation and all CsI included. These findings demonstrate that scientifically consistent augmentation data can meaningfully boost the ML model learning and robustness for planetary science MS data, particularly in sparse-data regimes. By addressing the challenge of limited training data, this work lays a critical foundation for developing reliable ML models to support science autonomy development in future planetary missions.

Our main goal is to support the science team with their analysis and decision-making process using data science and ML tools during operations of the MOMA instrument on Mars. As such, our work was built on a solid collaboration to fully address the science team’s needs and develop the most suitable tools while respecting mission operations constraints. Our multidisciplinary team composed of MS experts (to guide the science evaluation of the analysis tools), data science and ML experts (to develop the tools using computer science techniques), and software engineers (to develop the user interface for the tools) closely worked together from the problem statement, through the data collection and processing, to the model evaluation and deployment.

A key focus of this research was the systematic investigation of data augmentation to improve model generalization and robustness. Our findings provide concrete, data-driven recommendations for configuring NN in planetary mission contexts with scientifically validated augmentation data. Building on this research, future work will expand toward NN-based data augmentation methods, such as Generative Adversarial Networks (GANs; Goodfellow et al., 2020) and Variational Autoencoders (VAEs) to produce high-fidelity and scientifically-reliable mass spectra. GANs, for instance, use two NNs – one generating synthetic but realistic data points and the other distinguishing real from synthetic data points–to create highly realistic data, while VAEs learn latent representations that can be sampled to generate plausible new spectra. Although these techniques are computationally intense and more complex to implement, they hold significant potential to enrich training datasets, particularly in scenarios where data are sparse. Rigorous validation by domain experts will remain essential to ensure that any synthetically generated spectra preserve the science integrity required for MS data in planetary science applications.

Our team is also investigating the deployment of these ML prediction algorithms on high performance space computing (HPSC) boards from Microchip to test whether the MOMA flight software and the NN models can operate in parallel. This will include assessing inference performance from the NN models on newly collected data, as well as evaluating execution time, memory capabilities, and power consumption under flight-relevant conditions.

Beyond immediate applications to the MOMA instrument onboard ExoMars, our work directly benefits future missions such as Dragonfly, targeting Titan in the 2030s which will be equipped with DraMS a mass spectrometer instrument based on MOMA heritage, as well as missions targeting the Ocean Worlds Europa and Enceladus. Moreover, our multidisciplinary team (MS experts, Mars and Titan science experts, and computer scientists) validates the importance of collaboration between missions’ science team and data scientists in order to develop the most efficient approach for the analysis of MS data for space missions. This collaborative approach can be leveraged for future applications to other planetary instruments by identifying common challenges (such as data processing, decision-making support, anomaly/outlier detection, data prioritization) and producing general algorithms applicable to new instrument types.