AI Accelerates Universe and Planet Formation Simulations, Bringing Origins into Focus
The formation of our solar system and the universe itself can now be simulated in minutes rather than months, thanks to innovative machine learning techniques developed by CDS Senior Research Scientist Shirley Ho and her co-authors.
In two new preprints, Ho and her co-authors created AI models that accelerate complex simulations of cosmic and planetary evolution by up to four orders of magnitude. These models offer unprecedented insights into the conditions that shaped our universe and solar system.
“We want to understand how we got here,” Ho said. “That includes how the universe formed initially and what our planetary system was like when it first formed. How thick was the protoplanetary disk? Was it thinner? How many young stars were there?”
The first model is described in “Field-level Emulation of Cosmic Structure Formation with Cosmology and Redshift Dependence,” led by Max Planck postdoc Drew Jamieson. The model in this paper can rapidly generate simulations of the universe’s evolution for various cosmological parameters and initial conditions. This breakthrough allows researchers to explore a wide range of scenarios that were previously computationally infeasible.
Ho and her co-authors innovated by pre-training their model on a single universe setting before fine-tuning it for multiple scenarios, similar to how language models like ChatGPT are trained. “We pre-trained our model based on our previous model in which we can only predict one time-step of the Universe. This helps the model predict all other time steps,” Ho explained.
They also incorporated physics directly into the model. “Physics is at the heart of the design of this neural network,” Ho said. “Our loss function requires a specific relationship between time-dependent particle coordinates and the velocities.”
The second model, detailed in “Accelerating Giant Impact Simulations with Machine Learning,” focuses on planetary formation. Led by Princeton Astrophysics PhD student Caleb Lammers, this research uses machine learning to predict collisions between planets and their outcomes in nascent solar systems.
These advancements build on earlier work by Ho and her colleagues, including a 2021 paper published in the Proceedings of the National Academy of Sciences titled “A Bayesian neural network predicts the dissolution of compact planetary systems.” This earlier study marked the first application of machine learning to planetary dynamics, predicting in just 10,000 steps what would have taken a billion steps in traditional simulations.
“The fact you can accelerate these simulations makes the previously impossible possible,” Ho said. “Now you can compare the theoretical predictions of how planetary systems form to the observed planet statistics. You have a route to understand the beginning of planetary systems, including our own.”
Both projects emerged during the challenging times of the COVID-19 pandemic. “[This work] anchored us to the tininess of ourselves in a world of chaos at that point,” Ho reflected. “Without that, I would have really felt quite lost.”
As these models continue to evolve, they promise to unlock new insights into the fundamental questions of our cosmic origins, bringing us closer to understanding our place in the vast expanse of the universe.
Ho and her team are now working on even more ambitious projects, including building large foundation models for science — essentially creating an AI scientist that can learn across many fields of science.
By Stephen Thomas
