RocketSurrogate: A Surrogate Model Framework Using Machine Learning to Replace Costly Rocket Flight Simulations

This article introduces the open-source project RocketSurrogate, maintained by SuleimanGrape and released on GitHub on June 4, 2026 (link: https://github.com/SuleimanGrape/RocketSurrogate). This framework uses XGBoost and neural networks to build surrogate models, replacing RocketPy's costly numerical simulations to achieve fast inference and large-scale parameter sampling, with the core goal of solving efficiency issues in rocket simulations.

A single RocketPy simulation takes tens of seconds to over a minute, and training a high-quality surrogate model requires more than 50,000 samples, consuming significant computational resources. The project focuses on three core questions: 1. Can machine learning surrogate models accurately replace RocketPy for predicting flight results? 2. Which surrogate architecture balances accuracy and inference speed best? 3. How many simulations are needed to achieve acceptable accuracy?

The project implements a complete RocketPy 6-Degree-of-Freedom (6-DOF) simulation pipeline, including a timeout protection mechanism to ensure stable operation. Parameter sampling strategies include random sampling, Latin Hypercube Sampling (LHS), Sobol sequences, and balanced sampling, covering the design space. A two-stage validation is adopted: checking the physical feasibility of parameters before simulation, and verifying that results conform to physical laws after simulation.

XGBoost Surrogate: Supports native categorical features, includes feature engineering (length-diameter ratio, thrust-to-weight ratio, fin area ratio, etc.), an end-to-end simulation→training→evaluation process, and uses JSONL dataset format for easy expansion.

Neural Network Surrogate (under development): Plans to implement large-scale dataset generation with over 10,000 samples, knowledge distillation (integrating models into compact ones), and an active learning pipeline.

RocketPy simulations are costly (tens of seconds to minutes per run), while surrogate models exchange upfront data investment for millisecond-level inference. Application values include: 1. Fast exploration of design space; 2. Reducing the number of full simulations to save resources; 3. Supporting real-time optimization and interactive design; 4. Providing a demonstration for combining domain-specific simulators with ML.

Future Directions: 1. LLM-assisted design analysis (supports ROCm fine-tuning); 2. Active learning for intelligent parameter combination selection; 3. Knowledge distillation to compress models for edge deployment.

Project Structure: Modular design (src/rocket_sim for data generation, src/gbt for XGBoost surrogate, src/neural_surrogate for neural network surrogate, etc.).

Quick Start: Install dependencies → Generate synthetic data → End-to-end simulation training (see original project for specific commands).

RocketSurrogate demonstrates the potential of ML surrogate models in scientific computing and engineering simulation, achieving an order-of-magnitude speed improvement while maintaining accuracy. Its open-source implementation and documentation provide a reference template for fields requiring costly simulations such as fluid dynamics and materials science, and surrogate modeling is a direction worth exploring.