LLNL Open-Source Surrogate Model Course: A Complete Practical Path from Neural Networks to Bayesian Optimization

SEAMsurrogates, an open-source educational resource released by Lawrence Livermore National Laboratory (SEAM), covers four core technologies: neural networks, Gaussian processes, Bayesian optimization, and sensitivity analysis, with complete datasets and a 12-week learning path. Maintained by LLNL researchers like Jason Bernstein, it was released on GitHub in 2025 (link: https://github.com/llnl/SEAMsurrogates), and the documentation is available at https://seamsurrogates.readthedocs.io/.

In scientific computing and engineering design, expensive simulation tasks (such as fusion simulation and spacecraft reentry heat flux prediction) make traditional iterative optimization impractical. Surrogate models replace real simulations with low-cost approximations to improve efficiency, which are crucial in engineering optimization and uncertainty quantification. As a renowned U.S. national laboratory, LLNL has compiled its experience in applying surrogate models in fields like nuclear physics and climate simulation into the SEAMsurrogates open-source course.

The course follows a 12-week progressive learning path: the first 2 weeks cover deep learning fundamentals (PyTorch), week 3 focuses on surrogate model application practice, weeks 4-5 on Gaussian process regression (uncertainty estimation), weeks 6-7 on Bayesian optimization (intelligent sampling to find optimal solutions), week 8 on sensitivity analysis (parameter impact assessment), and the last 4 weeks on comprehensive projects. The core toolkit surmod includes four modules:

• Neural network module: FFNN workflow (preprocessing, training, validation);
• Gaussian process module: training and prediction, providing confidence intervals;
• Bayesian optimization module: acquisition functions (e.g., expected improvement), supporting high-dimensional constraints;
• Sensitivity analysis module: global methods to identify key parameters.

The course provides datasets from three domains:

1. JAG ICF dataset: Inertial confinement fusion research, nonlinear multi-physics coupling;
2. Drilling dataset: Groundwater flow benchmark problem, demonstrating parameterized simulation applications;
3. Hubble Telescope dataset: Thermal protection system temperature prediction, transient simulation scenarios. These datasets cover typical patterns like nonlinearity, high cost, and multi-parameter inputs, helping to accumulate experience with real engineering problems.

Surrogate models demonstrate value in multiple scenarios:

• Engineering design optimization: Reduce the number of simulations from thousands to tens/hundreds;
• Uncertainty quantification: Lower the cost of Monte Carlo sampling;
• Real-time decision support: Obtain approximate results in milliseconds. By open-sourcing this resource, LLNL provides a window for domestic researchers in related fields to learn cutting-edge technologies.

Installation: Install the surmod package via pip (automatically handles dependencies). Learning suggestions:

• Follow the 12-week course, complete code exercises, and do not skip the basics;
• Those with machine learning foundations can directly jump to topics of interest;
• The final project encourages applying the knowledge to your own research problems. Course materials include Jupyter Notebooks and annotated Python scripts; for questions, refer to the online documentation or example code.

SEAMsurrogates is an ideal open-source educational resource: it integrates frontline experience from top institutions, a complete theoretical system, rich practical data, and a clear learning path. It not only provides tools but also conveys the way of thinking to understand surrogate models. For engineers and researchers seeking surrogate model learning resources, it is a project worth in-depth study.