⚡ electrical engineering

Comparative Performance Analysis of Xenon, Krypton, and Argon Propellants for Hall Thrusters Using a Reproducible Machine-learning Framework

This paper presents a reproducible machine-learning framework using a harmonized dataset to demonstrate that while Xenon remains the superior propellant for Hall thrusters, Krypton offers a cost-effective compromise and Argon requires specific optimization, with XGBoost emerging as the most accurate model for predicting thrust performance.

Original authors: Nazmul Hasan Anik Chawdhury

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nazmul Hasan Anik Chawdhury

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Choosing the Right "Fuel" for Space Rockets

Imagine you are building a very special kind of car that doesn't run on gasoline but on electricity. This car is a Hall Thruster, a type of engine used by satellites to move around in space. Just like a car needs gas, this engine needs a "propellant" (a gas) to shoot out the back and push the satellite forward.

For a long time, the space industry has used Xenon as its fuel. It's like using premium, high-octane gasoline. It works incredibly well, but it is also incredibly expensive and hard to get.

This paper asks a simple question: Can we use cheaper, more common gases like Krypton or Argon instead, and still get the job done?

To answer this, the authors didn't just build a new engine in a lab. Instead, they acted like super-sleuths. They gathered 600 different "test drives" (data points) from hundreds of scientific papers, cleaned them up, and fed them into a Machine Learning computer program. Think of this program as a very smart, tireless mechanic who can look at all that data and say, "If you use Gas A at this speed, you'll get this much push. If you use Gas B, you'll get that much."

The "Smart Mechanic" (The Machine Learning Models)

The researchers tried five different types of "smart mechanics" (algorithms) to see which one could predict the engine's performance best. They were:

Random Forest (A committee of decision trees)
Gradient Boosting (A student who learns from its mistakes)
XGBoost (A super-charged, highly optimized student)
Neural Network (A computer brain that mimics the human mind)
Gaussian Process (A cautious predictor that always accounts for uncertainty)

The Winner: The XGBoost model was the clear champion. It predicted the engine's "thrust" (how hard it pushes) with amazing accuracy, getting it right about 98% of the time. It was like a mechanic who could guess exactly how fast your car would go just by looking at the gas type and the throttle setting.

The Showdown: Xenon vs. Krypton vs. Argon

Once the "smart mechanic" was trained, the researchers used it to compare the three gases. Here is what they found, using a simple analogy:

Xenon (The Luxury Sports Car):
- Performance: It is the heavy hitter. Because Xenon atoms are heavy, they carry a lot of momentum. When shot out of the engine, they provide the strongest push and the highest efficiency.
- The Catch: It is the most expensive fuel. It's like driving a Ferrari; it performs beautifully, but the gas costs a fortune.
Krypton (The Reliable Sedan):
- Performance: It's lighter than Xenon, so it doesn't push quite as hard. It's a bit less efficient, meaning you might need a little more energy to get the same result.
- The Benefit: It is much cheaper and easier to find. It's the "Goldilocks" option—not the best performer, but not the worst, and it saves you a lot of money.
Argon (The Economy Bike):
- Performance: It is the lightest of the three. Because it's so light, it's harder to get a strong push out of it. The engine has to work harder to ionize (charge) the gas, and the "exhaust" tends to spread out more, wasting energy.
- The Benefit: It is dirt cheap and everywhere. However, to make it work well, you have to tune the engine very carefully. If you don't, it's like trying to race a bicycle against a sports car; it just won't keep up without major modifications.

The "Cost vs. Performance" Scorecard

The paper didn't just look at who was fastest; it looked at who gave the best bang for the buck.

They created a score that balances how well the gas works against how much it costs.

Xenon wins on pure performance but loses on cost.
Argon wins on cost but loses on performance (unless the engine is perfectly redesigned for it).
Krypton sits right in the middle. The paper suggests that for many modern satellite missions, Krypton is the smartest compromise. It offers a good balance: you don't lose too much performance, but you save a massive amount of money.

Important Caveats (What the Paper Doesn't Say)

It is important to understand what this study is not doing:

It's not a physics experiment: The computer didn't simulate the tiny, chaotic dance of electrons and ions inside the engine (the "microphysics"). It simply learned the pattern between "Input A" (gas type, voltage) and "Output B" (thrust).
It's not a magic crystal ball: The model is very good at predicting results for conditions similar to the data it was trained on (interpolation). If you tried to use it for a completely new, futuristic engine design that has never been tested, it might get confused (extrapolation).
It's not a final verdict: The paper concludes that there is no single "best" gas for every situation. It provides a tool for engineers to make their own decisions based on their specific budget and mission needs.

The Bottom Line

This paper built a digital "test track" using data from the past to predict the future. It confirmed that while Xenon is still the king of performance, Krypton is a very strong, cost-effective challenger, and Argon is a budget-friendly option that requires careful engineering to work well. The "smart mechanic" (XGBoost) proved that we can use data to make smarter, cheaper choices for space travel without needing to build a new physical engine for every test.

Technical Summary: Comparative Performance Analysis of Xenon, Krypton, and Argon Propellants for Hall Thrusters

Problem Statement
Hall-effect thrusters (HETs) are critical for modern satellite missions, including orbit raising, station keeping, and deep-space exploration. While Xenon (Xe) has historically been the reference propellant due to its high atomic mass and low first ionization energy, supply constraints and high costs have driven interest in alternatives, specifically Krypton (Kr) and Argon (Ar). However, predicting the macroscopic performance of HETs across different propellants is challenging due to the complex, coupled physics of ionization, electron transport, and wall interactions. Purely analytical models struggle to capture these nuances, while experimental testing for every potential operating condition is resource-intensive. There is a need for a data-driven approach that can interpolate performance trends within a measured operating envelope to assist in propellant selection and mission trade-offs.

Methodology
The study employs a reproducible machine-learning (ML) framework to compare Xe, Kr, and Ar propellants. The core of the methodology involves:

Dataset Compilation: A harmonized dataset of 600 operating points was constructed from peer-reviewed journals, conference proceedings, and technical reports. Data was standardized to common units (Volts, Amperes, Watts, mg/s, mN, seconds) and screened to ensure recoverable input and output variables. The dataset covers a wide power range (0.1 kW to 72.5 kW) and includes various thruster geometries.
Feature Engineering: Input vectors included discharge voltage ( $V_d$ ), discharge current ( $I_d$ ), input power ( $P_{in}$ ), propellant mass flow rate ( $\dot{m}$ ), magnetic field indicators ( $B_r$ ), channel geometry, and intrinsic propellant properties (atomic mass, first ionization energy). One-hot encoding was used for propellant identity.
Model Evaluation: Five supervised regression algorithms were evaluated: Random Forest, Gradient Boosting, XGBoost, Multi-Layer Perceptron (MLP), and Gaussian Process Regression (GPR).
Validation Strategy: To ensure robustness, the study utilized multiple validation strategies:
- Random Split: An 80/20 split with 5-fold cross-validation to assess interpolation capabilities.
- Grouped Validation: "Leave-one-campaign-out" and "Leave-one-propellant-out" tests to estimate generalization to unseen hardware configurations and propellant types.
- Uncertainty Quantification: Bootstrap resampling for tree-based models and native posterior variance for GPR to establish predictive intervals.
Cost-Performance Index: A normalized thrust-per-cost index ( $C_T^*$ ) was developed, incorporating relative propellant costs (normalized to Argon) and predicted performance metrics to facilitate mission-level economic trade-offs.

Key Contributions
The paper claims six primary contributions:

A harmonized, literature-based dataset of 600 HET operating points covering Xe, Kr, and Ar.
A consistent feature-engineering statement for HET performance modeling.
A comparative analysis of five ML model families for predicting thrust, specific impulse ( $I_{sp}$ ), and efficiency.
Rigorous validation including grouped tests (campaign and propellant leave-outs) and uncertainty reporting.
A physically interpreted comparison of propellant performance trends derived from the trained models.
An explicitly defined cost-performance index with sensitivity analysis to cost assumptions.

Results

Model Performance: XGBoost emerged as the superior model for thrust prediction in the random-split evaluation, achieving an $R^2$ of 0.978, RMSE of 1.21 mN, and MAE of 0.86 mN. It outperformed Gradient Boosting, Random Forest, MLP, and GPR.
Generalization: Predictive accuracy decreased significantly in grouped validation scenarios. The "Leave-one-propellant-out" test yielded a thrust $R^2$ of 0.701, indicating that while models interpolate well within known data, extrapolating to entirely new propellant physics or hardware campaigns remains challenging.
Propellant Trends: The models reproduced expected physical trends:
- Xenon: Delivers the highest thrust and efficiency due to high atomic mass and favorable ionization characteristics.
- Krypton: Offers a practical compromise, providing intermediate thrust and efficiency with lower cost.
- Argon: Shows the lowest thrust and efficiency under standard conditions due to lower atomic mass and higher ionization energy, requiring careful thruster optimization to mitigate utilization losses.
Feature Importance: Input power and atomic mass were identified as the strongest predictors of thrust in the XGBoost model.
Cost-Performance Trade-off: The normalized index ( $C_T^*$ ) suggests that while Xenon is the performance leader, Krypton offers a robust compromise. Argon ranks highest in cost-normalized terms but carries significant technical risks regarding utilization and plume divergence unless the thruster is specifically optimized for it.

Significance and Claims
The authors position this work as a data-driven engineering screening tool rather than a substitute for plasma-resolved simulations or physical experiments. The study explicitly states that the ML models capture macroscopic input-output trends and do not resolve underlying plasma microphysics (e.g., electron energy distribution functions or ionization cross-sections).

The significance of the framework lies in its reproducibility and its ability to link operating conditions, propellant properties, prediction uncertainty, and mission-level economic assumptions. The paper concludes that no single propellant is universally superior; rather, the choice depends on the specific mission constraints regarding cost, availability, and acceptable efficiency penalties. The framework provides a structured method for evaluating these trade-offs, particularly highlighting Krypton as a viable alternative to Xenon when system economics are a priority, and Argon as a cost-effective option contingent upon thruster-level optimization.