Machine learning is a sub-discipline of computer science that aims to develop systems that can learn from data in order to support decision making processes. While several types of ML algorithms exist, the two main categories are supervised learning and unsupervised learning (Figure 1). In supervised learning, input data is given with labels: X usually represents an m by p matrix of input data with p features (equivalently referred to as variables, predictors or dimensions) and m samples or observations, and y represents a vector of class labels for the samples that the algorithms will learn to predict. The main goal is to use algorithms to optimize mathematical (statistical) models of the p features to best fit the class label y. For supervised learning optimization, it is important to include cross-validation and, when feature selection is used, nested cross-validation to avoid overfitting (Parvandeh et al., 2020). These models are then used to predict an output given new data previously unseen by the model. Algorithms learn by minimizing the error measured between the algorithms’ results and the correct outputs (given labels y). Common applications of supervised ML are for prediction–making a prediction of the output label for new unseen data, and information–helping to understand the relationship between inputs and output(s). In contrast, unsupervised learning input samples do not include external labels, y, for making predictions, and, thus, do not require cross-validation. The main goals of unsupervised learning are to explore the features and samples in the data matrix X and study the intrinsic structure of the data. Unsupervised ML is often used for clustering–grouping the data in similar groups or by similar features–and for dimensionality reduction–simplifying the datasets by reducing the number of features (dimensions).

During the advancement of a data science or ML project, it is essential to keep in mind that 1) the choice of the approach depends on the goal of the project, and 2) the goal(s) must be realistic and achievable. Moreover, the approach needs to be defined based on the structure and the properties of the data, which EDA helps the scientists understand. The problem to solve, and the available resources like data and domain knowledge will change the strategy to apply. Moreover, the science experts (in our case, mass spectrometry scientists) and data scientists need to work in close collaboration to correctly implement tools to achieve the main science goal(s), while respecting the project or mission constraints. In this work, we introduce a data science and ML pipeline for EDA of ocean worlds analog spectra, keeping in mind that similar techniques could be applied to other planetary instruments’ datasets.

Here we use laboratory analog datasets from Theiling (2021), as well as new data from ocean worlds seawater analogs and associated CO₂ that interacted with the seawater (Figure 2). These Enceladus- and Europa-relevant laboratory analog ocean world seawaters were previously analyzed to refine the current carbon and oxygen isotopic fractionation models for seawaters of various composition and pH (Theiling, 2021). In the present study, these data are used to understand the potential of CO₂ spectra to cluster by seawater environment. These seawaters are composed of varying ionic strengths and combinations of NaCl, MgCl₂, NaHCO₃, MgSO₄, and Na₂SO₄. CO₂ mixed with helium was injected in varying concentrations (0%, 0.3%, 1%, and 2%) to study the isotopic fractionation that occurs as each solution interacts with the CO₂. Each experiment was prepared and analyzed in triplicate. Experiments were equilibrated over varying times: 24, 48, 72, and 168 h. The standard equilibration time used after initial experimentation was 168 h. Isotope ratios (δ¹³C and δ¹⁸O) were analyzed using a trace gas analyzer (Gasbench II) coupled to an isotope ratio mass spectrometer (IRMS: Thermo Finnigan Delta V Advantage). Theiling (2021) demonstrates that the CO₂ concentration and pH affect the isotopic fractionation, which would affect isotope ratios measured during missions to Enceladus or Europa differently. In our present study, we use this CO₂ IRMS dataset to develop EDA methods, automated quality checks, processing, filtering, data analysis, and ML pipelines for predictions of seawater composition and identification of biosignatures. For EDA methods detailed in the current study, these proof-of-concept tools include basic data analysis methods and unsupervised ML techniques for detection and visualization of the data structure. Supervised ML techniques to learn underlying complex models and predict the composition of new, unseen data are the subject of future work. The driving goal of these algorithms is to develop an analytical software tool that could be used by mission scientists as a rapid data analysis method for returned data. A longer-term goal is to deploy further analytical ML tools onboard a spacecraft for automated real-time data analysis and decision-making.

Data from each experiment is composed of several measurements of volatiles over time in the headspace of a 12 mL vial containing 0.5 mL of an analog seawater. Experiments consist of repeated sampling using a double walled needle that exchanges sample gas with inert helium. These volatile measurements are then transferred to the IRMS through an 8-way valve, which effectively produces 10 discrete “sample” peaks (Figure 3). The data collection strategy was organized around several factors, with variations of each one as described in Table 1.

Data from each experiment is composed of raw chromatograms (x-axis: time (s), y-axis: voltage (mV)) and extracted data in tables representing spectral features such as isotope ratios (for CO₂), voltages (mV), and times (s), calculated features such as the fractionation values of these ratios (δ¹³C, δ¹⁸O), as well as experiment metadata such as CO₂ preparation concentrations, compound information, etc. Before training an ML algorithm to classify biotic versus abiotic compounds for instance, or to classify between the amount of injected CO₂, we need to understand the dataset and process it into high-quality input data for further data science and ML algorithm investigation/development. The first step consists of pre-processing the data from raw experimental data using automated quality analysis/quality control (QA/QC) designed using expert knowledge of our dataset, replicating the checks usually performed by science operations teams.

After performing QA/QC checks, we perform Exploratory Data Analysis (EDA)—a critical process of performing initial investigations on data in order to discover patterns, find anomalies such as outliers, and test initial hypothesis and assumptions with some statistics and visual representations. EDA is a tool used to better understand the dataset and prepare it as well as possible for further ML algorithms implementation. Indeed, EDA is an initial and fundamental step to learn patterns from your data that could then be leveraged to create precision algorithms that focus on onboard capabilities to support science investigations.

As mentioned above, unsupervised learning algorithms are used to explore and analyze the structure of the data to identify patterns and similarities (clustering) or to simplify the dataset (dimensionality reduction) without being influenced by an assigned class label. Unsupervised learning algorithms are applicable when the phenomenon driving the data is unknown. Rather than making predictions about unknown data, the unsupervised algorithm provides insights on the relationships within the data, allowing us to see patterns otherwise not recognizable by human investigators. A first step is typically to use dimensionality reduction algorithms to represent the high-dimensional input data into a low dimensional space that retains the inter-variable relationships of the original data. This first step aims at getting a better understanding and interpretation of the data, and is often used as a precursor step to use clustering techniques to find unbiased patterns and similarities that group data into clusters of similar and dissimilar information.

We use EDA techniques such as basic statistics (average, quartiles, minimum, and maximum values) and visualization plots like boxplots and pairplots in an iterative process for the verification of the QA/QC steps. After each implementation of a new QC, pairplots and boxplots are generated to verify the presence of potential outliers representing poor quality laboratory data that should be filtered out, which are cross-checked with more laborious QA/QC procedures typical for evaluating similar data types, including visual inspection of chromatograms and filtering, sorting, and calibration procedures performed in spreadsheets. Unlike analysis of natural samples, in which the purpose is to characterize unknown materials, the purpose of the ocean worlds experimental dataset we use is to understand similarities and differences in output under highly-constrained conditions. Therefore, we must understand how variable data quality might interfere with data from the experimental conditions (i.e., reduce the probability of false positives/negatives, aid in instrument diagnosis).

Figure 9 illustrates five features [“PeakNr” (peak number), “Rt” (retention time), d13C12C, d18O16O, and Ampl44] from unfiltered data compared to those that underwent a first quality check implementation for the 5 reference peaks. We first focus on the 5 reference peaks as they offer clear readability of the outliers. We quickly observe that the number of peaks is reduced from 32 to 16 (last reference peak is peak #16), and the values range of d13C12C and d18O16O is rescaled. This suggests that our EDA process could assist in identifying poor quality data, and also demonstrates the improvement of the dataset using established QC checks on overall data. This is also important to note because a user could use additional EDA or ML on raw, QA/QC checked, or QA/QC checked and calibrated data. These pairplots therefore demonstrate that additional processing of data will improve results. Thus, EDA techniques could be used on raw data to understand structure and potential QA/QC flags, and then again on more processed data to evaluate scientific questions.

Correlation matrices depict the correlation between all possible pairs of values in the table. This is a powerful tool to summarize datasets with many features and to identify patterns in the given data. In Figure 10, we represent a correlation matrix for 3 different sets of features: 1) all the 41 features, 2) a limited set of 25 features, and 3) an even more limited set of 14 features. Limited features in Figure 4B are selected based on characteristics typically used to evaluate chromatograms, and these features were further down-selected in 4c based on correlations demonstrated in 4b. The choice of which features to remove is based on the high positive correlations represented by the green squares and a correlation value equal to 1. For instance, for a given isotope with mass X, “rIntensityX”, “IntensityX”, and “AmplX” contain the same information and are considered redundant in terms of their information content, although their absolute values may be unique. Removing highly correlated variables is needed to prepare the dataset for ML, streamlining processing and improving interpretability of the model.

Moreover, down-selecting features improves dimensionality reduction techniques. In the case of applying PCA (i.e., a dimensionality reduction tool that helps identify patterns in the dataset based on the correlations between features), removing highly correlated features enhances the output reduced set. The mathematical process of PCA aims at finding the directions of maximum variance in the high-dimensional space and at projecting it a lower dimension space. By keeping strongly correlated features before performing PCA, the contribution of their common underlying factor increases and can have substantial effect on PCA results. Even if PCA is not a feature selection method, it is common to consider the ranking of the features in the explained variance and only retain the most important features that explain most of the variance.

We analyze the role of each feature using 3 different sets of input features for PCA application (Figure 11): 10 features (Figures 11A, D), 5 features (Figures 11B, E), and 2 features (Figures 11C, F). When plotting the PCA results we add vectors representing the direction, orientation, and amplitude of each input feature. Figure 12 illustrates the high correlation between some input features such as Start/End (representing the timing of the sample or standard peak), d13C12C/d18O16O (representing the ratio of CO₂ isotopologues measured in the experiment). Using these PCA results, we are then able to reduce the input features to best represent the dataset while preserving the most information.

Similar to PCA, t-SNE and UMAP representations in Figure 7, the linear patterns emerging in Figures 12A–C are due to repeated measures of each sample peak. Therefore, to remove the bias in continued data science and ML techniques, we average the repeated measures, resulting in more clear cluster structure (Figures 12D–F). In the raw data, each experiment consists of CO₂ equilibrated with an ocean world analog seawater, and repeated sampling by the IRMS results in 9 sample peaks used for further analysis (Figure 3). As a simple first transformation, we create a separate dataset in which the 9 sample peak values are averaged into a single value, and all potential features including a time component (e.g., Rt, Start, End, etc.) are removed. This averaging of the repeated measures results in a dataset, that is, more conducive to further analysis and ML.

We analyze the differences of the repeated measures dataset (only features with correlation lower than 0.9 have been used in these representations) and the time-averaged dataset. We note that PCA, t-SNE and UMAP algorithms used as dimensionality reduction techniques are able to simplify the clusters when using the mean-reduced data versus the repeated-measures data. By limiting the number of repeated measures per experiment (by averaging) and by limiting the numbers of features (only the ones with correlation values <0.9), these basic dimensionality reduction algorithms demonstrate they are better at clustering than the raw data.

Solar system missions not only present operational challenges but also technological ones. Indeed, the harsh environment the spacecraft are exposed to, such as extreme temperature gradients and radiation, constantly damage the electronics components, limiting the diversity of hardware usable on spacecraft. An additional challenge therefore is accommodating the computational requirements for onboard data science and ML processing. These computing requirements pose a difficult problem for the application to space missions, which are extremely limited in power consumption, communication bandwidth, and hardware equipment. The majority of NASA missions require “Radiation Hardened” (RH) electronics, including everything from power supplies to general purpose central processing units (CPUs), or requirements for certain electronics to be house in a radiation hardened vault. These requirements limit the available power and data volume, creating a widespread need for technological development in RH processing capability. As an example, the Curiosity rover, the Perseverance rover, and the Mars Reconnaissance Orbiter (MRO) are all equipped with the same RH CPU (BAR Systems RAD750 processor) that was developed in the 1990s. The demonstrated flight heritage (TRL 9) and reliability of the RAD 750 makes it a standard in space missions, despite many generations behind the state-of-the-art in processors. The emergence of FPGAs (Field-Programmable Gate Array, i.e., integrated circuit designed to be configured by a customer) with DSP (Digital Signal Processors) or simple integrated neural processors shows promise that sufficient AI and ML application hardware will soon be developed for space flight applications. However, the mission design requires an exploration of tradeoffs between runtime and accuracy to offer tools that answer science objectives while respecting operational constraints. ML evolves very quickly and the latest sophisticated models cannot fit in the severe resource constraints. To date, computer science tools that have been deployed in space have low computational demands. One of the next steps is to monitor the power, memory, and computation time for the methods mentioned in this paper. During the development of these algorithms, we kept in mind the needs of performance and resource-optimization due to high limitations for planetary missions.

Moreover, ML applications require large amounts of data in order to properly train and tune the developed algorithms. In this work, we use data collected in a laboratory over 2 years. The total dataset was composed of 1,086 input spectra, reduced to 587 spectra after the quality check process. While more data are being collected in the laboratory, this process can take many years; therefore. Data generation techniques to enhance the input dataset, such as synthetic data creation or data augmentation methods (Maharana et al., 2022; Mikolajczyk et al., 2018) could be investigated in future research. This work is a proof-of-concept demonstrating a powerful data science pipeline that can support scientific research in a multitude of ways—from implementing automated checks on data quality and reproducibility, to understanding emergent patterns in datasets that may be outside typical uses of a dataset. By introducing these procedures here, we endeavor to inspire development of science- and mission-enhancing tools, including deployment of onboard analysis. We acknowledge, however, that data from flight instruments are often significantly different from commercial instruments, therefore models developed on commercial instrument data will need to be updated to be implemented on actual flight instrument data.

The investigation and development of intelligent algorithms to support space mission operations, using data science, basic thresholds methods, AI, or ML require the mission and instrument design to accommodate this innovative implementation. One of the key points is the creation of the dataset used to train and test the algorithms. The dataset must be well defined from the beginning of the mission concept in order to support effective and accurate algorithms to answer the specific requirements of the mission (operational, scientific, and engineering). Instrument scientists must establish precise goals and a coherent and clear dataset strategy creation. For this research, data scientists had to preprocess the dataset several times to remove unsuitable data and to reformat data collected in inconsistent ways by multiple users. Therefore, thinking about ML or data science applications from an early stage in the mission design will drive the data collection plan and greatly enhance the possibilities of data-driven algorithmic investigations.

The increasing presence of AI and ML is also accompanied with rising concerns, especially in domains like space exploration. The deployment of onboard intelligence techniques is still avant-garde and consequently high-risk. Every new technology or capability confronts a trade-off between the new benefits added to the instrument and deliverables, and the potential additional risks from its implementation. ML techniques must be fully tested and proven in order to demonstrate ML reliability. Moreover, it is worth mentioning that traditional algorithmic approaches such as threshold-based algorithms should be first considered if they enable the same science benefits while being more easily implementable onboard a mission than more opaque data science techniques. Algorithms must be designed in order to provide interpretability and limit this “black-box” effect.

Another important characteristic in the design and conception of AI and ML methods is the choice of the performance metrics based on the interest of the stakeholders. In our research, ML metrics such as accuracy, recall, and precision are being used in combination with scientists’ interpretations. The communication between the fields of data science and mass spectrometry is crucial to fully understand the needs of the science team as well as the limitations they face, and will be instrumental in promoting more widespread use and development of these tools among the scientific community.

Slingerlan and Perry (2022) provides a framework for developing trust in AI as further applications will require careful considerations of the training and the deployability of software and hardware in more realistic and therefore constrained environments. Recent research considered the development of a “Trust Readiness Level”, similar to the “Technology Readiness Level” (TRL) used in the spaceflight engineering field for hardware ability evaluation (Da Poian et al., 2022). Our research aims at designing a scale system to assess the maturity level of the technology using AI and ML tools. Similarly to the engineering scale, this trusting system will follow a progressing path that tests these methods on laboratory data in normal conditions, and then under more realistic flight mission conditions and constraints before being actually used onboard a mission. This Trust Readiness Level scale will not only be dependent of the target of interest but also on the types of instruments. Because mass spectrometers going to Mars will not have the same constraints and requirements as the ones going further out in our Solar System (to Titan onboard Dragonfly, to Europa onboard the Europa Clipper mission), this trust readiness level scale will be dependent of the target of interest for the mission. It will also be dependent on the type of instruments under study while studying the same target of interest, as for instance mass spectrometers and magnetometers will not hold the same requirements and constraints and will though have different trust demands. Developing this trust readiness level will require a systematic methodology, consistency, and adjustment on a case-by-case basis.