Optimization methods for hydrodynamic simulations using machine learning surrogate models and LLMs
Please login to view abstract download link
New capabilities in machine learning (ML) and high performance computing are creating exciting opportunities for inverse design optimization of physics simulations. ML surrogate models, which can be used to relate design parameters to specific objective metrics or fields (e.g., density, pressure, etc.), provide an efficient means for optimizing a set of design parameters once the surrogate model is trained on simulation data. These types of surrogate models may use various modelling methods, from Gaussian process regression to generative models such as GANs and diffusion models. For example, Bayesian optimization uses GPR surrogate models as a means of iteratively exploring the design space [1]. We develop a framework that uses a large language model (LLM) in conjunction with ML surrogate models for the inverse design optimization of hydrodynamic simulations. LLMs provide new and unique capabilities for automating the design and optimization process [2]. As an example, a LLM can be used to efficiently explore a design parameter space using forward evaluations of a surrogate model to determine the optimal set of input design parameters for minimizing a given objective metric. A LLM can also be used to run new simulations that are informed by the surrogate model and train a new updated surrogate model. One major advantage of using a LLM for this purpose is that the LLM can intelligently produce a sophisticated multi-objective function based on text prompting from a human user to achieve a desired design outcome. A vision LLM can also use physical fields and plots to create this multi-objective function. This can bypass the need to develop a mathematical objective function that is determined in advance of the optimization, which is necessary for a traditional optimization method. We present results that use simulations of an inertial confinement fusion capsule implosion to demonstrate this framework. This work was performed under the auspices of the U.S. DOE by Lawrence Livermore National Laboratory (LLNL) under Contract DE-AC52-07NA27344 and was supported by the ASC Multi-Agent Design Assistant project at LLNL. IM release number: LLNL-ABS-2014719
