Balancing Variety and Sample Size: Optimal Parameter Sampling for Ariel Target Selection

Targeted astrophysical surveys are limited by the amount of telescope time available, which makes it impossible to observe every single object of interest. In order to maximize the scientific return, we need a well thought strategy for selecting the observational targets, in our case exoplanets. This study evaluates various strategies for selecting exoplanet targets within limited observation windows, focusing specifically on the selection of exoplanet targets for Tier 2 transit spectroscopy with ESA’s upcoming Ariel mission. We define three distinct selection criteria – sample size, variance, and leverage – and translate them into objective functions compatible with modern optimization algorithms. Specifically, we test five heuristics for maximizing sample leverage: leverage greedy, simulated annealing, K-means clustering, regular classes, and quantile classes. The performance of these methods is demonstrated through three practical exercises across one, two, and three parameters of diversity. Each criterion represents a unique trade-off between sample size, diversity, and total observation time. While a time-greedy approach maximizes the quantity of planets, it fails to capture diversity. Conversely, variance-greedy selection prioritizes diversity but introduces significant drawbacks: it oversamples rare cases and undersamples typical planets, ultimately reducing the total number of targets observed. Leverage-based selections emerge as the most effective middle ground, successfully balancing sample diversity with a robust sample size. This work supports the broader community effort to ensure that Ariel delivers the most diverse and scientifically valuable sample of exoplanet atmospheres within mission limits.

💡 Research Summary

The paper addresses the problem of selecting a scientifically optimal subset of exoplanet targets for ESA’s upcoming Ariel mission, given a hard constraint on total observing time. The authors formalize three objective functions: (1) sample size (maximizing the number of planets, N), (2) sample variance (maximizing the variance V of a chosen planetary parameter p), and (3) leverage (L = √(N·V), or equivalently L² = N·V), which balances size and diversity. They argue that leverage directly controls the uncertainty on linear population‑trend slopes (σ_m ∝ 1/L), making it the most scientifically relevant metric for Ariel’s population‑level goals.

Using the Ariel Mission Candidate Sample (MCS) as of August 2025, they restrict the pool to 1 342 planets that require ≤20 transits for Tier 2 signal‑to‑noise. The total observing budget is set to three years (≈1 095 days), and the required time for each planet is taken as three times the transit duration. The authors test five heuristic selection strategies: (i) regular‑class binning, (ii) quantile‑class binning, (iii) K‑means clustering, (iv) simulated annealing, and (v) a newly proposed “leverage‑greedy” algorithm. A random selection is also used as a baseline.

The leverage‑greedy method treats the problem as submodular maximization under a knapsack constraint. At each iteration it computes the marginal gain per unit time for every feasible planet: Δf_L/Δt = (p_i − \bar p_S)² / t_i, where \bar p_S is the current mean of the selected set. The planet with the highest ratio is added, and the process repeats until the time budget is exhausted. This greedy rule is provably within a factor (1 − 1/e) of the optimal submodular solution.

Experiments are performed in one‑, two‑, and three‑dimensional parameter spaces (R_p alone; R_p and T_p; R_p, T_p, and T_s). Across all cases, leverage‑greedy consistently yields the highest leverage value, achieving a good compromise between the number of selected planets and the spread of their parameters. The time‑greedy approach maximizes N but produces a low variance, resulting in a small L. The variance‑greedy approach maximizes V but selects far fewer planets, also yielding a low L. K‑means and class‑based methods attain intermediate L values, while simulated annealing offers only marginal improvements at a higher computational cost. Random selection performs the worst.

The authors discuss the practical implications: leverage‑greedy is simple to implement, computationally cheap, and theoretically sound, making it suitable for real‑time mission planning or dynamic schedule adjustments. Limitations include the need to update the sample mean after each addition and the fact that the study only considers a single scalar objective (leverage) rather than a multi‑objective framework that might incorporate additional scientific priorities (e.g., atmospheric composition diversity).

In conclusion, the paper demonstrates that optimizing the leverage metric via a greedy submodular algorithm provides the most balanced and scientifically valuable target list for Ariel’s Tier 2 survey. This approach offers a generalizable template for other exoplanet missions facing similar trade‑offs between sample size, diversity, and limited observing resources.