A research team has developed an advanced approach for predicting thrust in cold gas thrusters using a feedforward neural network optimized through stochastic gradient descent (SGD). By implementing a correlation matrix-based feature selection method to reduce parameter redundancy, the researchers significantly improved both the accuracy of thrust estimations and the convergence speed of the predictive model. The study, conducted to address the growing demand for precise propulsion performance in modern small satellite missions, demonstrates that SGD consistently outperforms the Adam optimizer in this specific engineering context, achieving lower mean square error (MSE) values and faster learning rates over the course of training iterations.
Cold gas thrusters represent a critical propulsion solution for the burgeoning “new space” era, particularly for platforms such as CubeSats and nanosatellites. These spacecraft require propulsion systems that are energy-efficient, compact, and affordable. Unlike chemical propulsion systems, which rely on combustion, cold gas thrusters function by expelling pressurized, inert, non-toxic propellants like nitrogen or carbon dioxide through a nozzle. This design simplicity ensures high reliability for essential spacecraft operations, including station-keeping, collision avoidance, and precision attitude control. As adoption rates rise within the global satellite community, the need for robust methods to predict their thrust output under diverse operational conditions has become increasingly apparent.
Traditional analytical models often fail to capture the complex, nonlinear interactions between interdependent parameters like nozzle geometry, storage pressure, and mass flow rate. These oversimplified mathematical models can limit the precision required for high-stakes mission planning. Recent advancements in machine learning, however, offer a transformative path forward. By leveraging large datasets, neural networks excel at uncovering the nonlinear relationships that define thruster performance. This capability allows for continuous refinement of predictions through sensor inputs, offering a level of adaptability that is particularly valuable in the unpredictable and dynamic environments of space missions.
The current study derived its training data from a comprehensive experimental dataset encompassing 2300 chemical substances, including 300 inorganic and 2000 organic propellants. Each substance underwent a rigorous filtering process to ensure practical feasibility for space-based applications. Researchers established six key criteria for inclusion, such as compatibility with an operational temperature range of 20 to 100 °C, phase stability between 0 and 40 °C, and an upper vapor pressure limit of 1 atm at 20 °C. These constraints were designed to filter out radioactive or economically infeasible materials while ensuring that the data used for model training reflected both the practical constraints of propellant handling and the demands of high-fidelity propulsion engineering.
To optimize the performance of the neural network, the research team implemented a parameter selection method based on a correlation matrix. Statistical analysis via the Pearson correlation coefficient was employed to quantify linear dependencies between input features and identify redundant variables. Notably, the study revealed a correlation of 1 between nozzle length and throat diameter, indicating that including both variables would introduce multicollinearity rather than additional information. By excluding the throat diameter and retaining nozzle length, the researchers simplified the model architecture while preserving the governing physical behavior derived from classical nozzle design theory, thereby enhancing the convergence efficiency of the neural network.
The final architecture of the proposed model is a feedforward neural network consisting of an input layer, two hidden layers, and a single-neuron output layer. The input layer processes five selected features, while the first hidden layer incorporates 50 neurons and the second contains 20 neurons. Both hidden layers employ the rectified linear unit (ReLU) activation function to introduce the non-linearity necessary for complex feature extraction. The output layer uses a linear activation function to predict the average thrust. Weights are initialized using the Xavier initialization technique to facilitate stable gradient propagation, ensuring the model maintains stability during the learning process.
A core component of the study involved a systematic comparison between two primary optimization algorithms: stochastic gradient descent (SGD) and Adam. The researchers tuned hyperparameters for both methods to ensure a fair assessment. SGD updates model parameters by computing gradients from individual training samples, a process that introduces noise that can help the model escape suboptimal local minima. In contrast, Adam combines momentum and RMSProp techniques to leverage smoothed gradient estimates. While Adam is often effective in high-dimensional or sparse-data scenarios, its reliance on moving averages resulted in overshooting and oscillations within the cold gas thruster modeling context.
Quantitative simulation results demonstrated the superiority of the SGD algorithm. Furthermore, monitoring the R² score, which measures the proportion of variance explained by the independent variables, showed that SGD achieved faster learning and reached a higher level of accuracy more rapidly. At 50 epochs, SGD exhibited a swifter increase in R² scores compared to the Adam method, confirming its efficiency for quick training in mission-critical applications. These findings underscore the importance of aligning optimization techniques with the specific gradient behaviors and data characteristics of the target application.
To ensure the robustness of these findings, the model was evaluated using both holdout and K-fold cross-validation methods. The holdout method, which split the data into 70% training and 30% testing/validation sets, offered a straightforward evaluation technique efficient for large datasets. A supplementary comparison confirmed that while the 5-fold cross-validation approach produced slightly higher accuracy, the holdout method provided highly comparable and reliable results within a significantly shorter computational timeframe. This validation supports the practical utility of the proposed neural network framework for ground-based development environments where rapid iterations are required.
The neural network model was also compared against alternative meta-modeling techniques, including decision trees, logistic regression, and support vector machines. The decision tree model returned an MSE of 0.093 and an R² score of 0.84, while the proposed neural network with SGD achieved an MSE of 0.011 and an R² score of 0.94 at epoch 200. These results highlight the neural network’s enhanced ability to capture the complex, nonlinear relationships in the experimental data. By demonstrating lower MSE and higher R² values, the study establishes the neural network model as a highly effective tool for thrust prediction in cold gas systems.
The potential applications of this model extend beyond laboratory-based thrust prediction. By providing a robust, data-driven approach, the framework can be seamlessly integrated into spacecraft propulsion system design workflows, aiding engineers in early-stage thruster selection and enabling more precise mission planning. The researchers suggest that the precision and adaptability of the model make it a valuable resource for enhancing decision-making in aerospace engineering. Future research is expected to focus on expanding the model to include hybrid or chemical propulsion systems and integrating real-time sensor feedback for onboard operational implementation.
This study concludes that while adaptive optimization methods remain powerful in various domains, the simplicity and noise-driven stochasticity of SGD offer a significant advantage for problems with well-bounded parameter spaces such as cold gas thruster modeling. The combination of optimized parameter selection and an efficient neural network structure provides a reliable foundation for predicting thrust under dynamic conditions. As satellite missions continue to demand higher levels of performance and efficiency, data-driven tools like the proposed neural network model will become increasingly vital for ensuring mission success and maximizing the capabilities of future space-based platforms.
Beyond mathematical modeling, the physical architecture of cold gas thruster systems relies on clean, regulated propellant transport from a pressurized tank through feed lines and dedicated filters. High-pressure and low-pressure regulators step down container pressures before propellant reaches a solenoid valve, which activates expulsion into a converging-diverging nozzle where expansion accelerates gas to produce thrust. In structuring the feature selection process, classical nozzle theory dictated that conical nozzle length depends directly on exit diameter, throat diameter, and divergence half-angle. Because these geometric variables share an inherent mathematical link under fixed flow expansion, retaining both length and throat diameter would introduce redundant physical information. Rather than employing principal component analysis, which obscures the physical interpretability of individual parameters, or shrinkage techniques sensitive to feature scaling, the authors utilized correlation matrices to retain transparent, physically meaningful inputs.
Rigorous data pre-processing and hyperparameter tuning also underpinned the network’s predictive stability. Across the screened propellant substances, experimental conditions were standardized at 20 degrees Celsius and 1 atmosphere, while numerical missing entries were replaced using median imputation. Anomalous data points identified by absolute Z-scores greater than 3 were capped to prevent distorted optimization trajectories. During optimizer configuration, learning rates were systematically tuned against validation loss, selecting 0.1 for stochastic gradient descent from a test range spanning 0.0001 to 0.1, alongside exponential decay parameters of 0.9 and 0.999 for the Adam optimizer. In quantitative evaluations across 100 epochs, an eighty-twenty holdout split attained a mean square error of 0.015 and 98.4 percent accuracy, closely matching five-fold cross-validation results of 0.014 error and 98.9 percent accuracy. Scatter plots tracking predicted against measured thrust values across the complete operational envelope further verified consistent precision without systematic bias at low or high thrust thresholds.
Subject of Research: Optimization of neural network performance for thrust prediction in cold gas propulsion systems.
Article Title: Thrust prediction for cold gas thrusters using an optimized neural network approach
Article References: Farhid, M., Aghdam, M. R. G., & Shameli, M. (2026). Thrust prediction for cold gas thrusters using an optimized neural network approach. Journal of Electric Propulsion, 5(1), Article 23. https://doi.org/10.1007/s44205-026-00200-1
Image Credits: AI Generated
DOI: 10.1007/s44205-026-00200-1
Keywords: Neural Networks, Cold Gas Thrusters, Space Propulsion, Machine Learning, Stochastic Gradient Descent, Adam Optimizer, Aerospace Engineering, CubeSat Technology, Thrust Prediction, Correlation Analysis, MATLAB, Nonlinear Dynamics, Satellite Manoeuvring, Feature Selection, Performance Optimization
Cite Scienmag News
Cassandra Pierce. (November 19, 2026). Stochastic Gradient Descent Improves Cold Gas Thruster Thrust Prediction Accuracy. Scienmag. https://scienmag.com/stochastic-gradient-descent-improves-cold-gas-thruster-thrust-prediction-accuracy/
Cassandra Pierce. "Stochastic Gradient Descent Improves Cold Gas Thruster Thrust Prediction Accuracy." Scienmag, 19 November 2026, https://scienmag.com/stochastic-gradient-descent-improves-cold-gas-thruster-thrust-prediction-accuracy/. Accessed 19 November 2026.
Cassandra Pierce. "Stochastic Gradient Descent Improves Cold Gas Thruster Thrust Prediction Accuracy." Scienmag. November 19, 2026. https://scienmag.com/stochastic-gradient-descent-improves-cold-gas-thruster-thrust-prediction-accuracy/








