Presentation Information
[Poster 1-12]Evolutionary code search for numerical simulation development
*Dennis G. Wilson1 (1. Sakana AI (Japan))
Digital twins depend on numerical models that are accurate, stable, and adaptable, yet developing and maintaining these models remains labor-intensive. We investigate whether evolutionary code search can automate part of this process by applying ShinkaEvolve, an open-source framework that uses large language models as mutation operators within an evolutionary search, to numerical simulation development.
We formulate simulation development as program evolution from observed trajectories. Candidate programs receive an initial state and generate future states autoregressively, allowing the search to modify executable simulation code rather than fit a fixed neural architecture. We benchmark the resulting programs on several spatiotemporal physical systems using a common rollout interface. Evaluation separates short-horizon prediction error from long-horizon stability and physical behavior.
Initial experiments establish the end-to-end feasibility of using ShinkaEvolve to generate, execute, and iteratively improve numerical simulation programs under a limited evaluation budget. The evolved artifacts remain explicit source code, making them directly inspectable and potentially easier to adapt than opaque learned surrogates. At the same time, the benchmark exposes practical challenges for evolutionary simulation development, including computationally expensive fitness evaluation, unstable rollouts, and the need for objectives that reward meaningful temporal evolution rather than persistence.
These results position evolutionary code search as a promising tool for constructing and updating numerical components of digital twins, including future ocean and Earth-system applications where models must evolve alongside observations and operational requirements.
We formulate simulation development as program evolution from observed trajectories. Candidate programs receive an initial state and generate future states autoregressively, allowing the search to modify executable simulation code rather than fit a fixed neural architecture. We benchmark the resulting programs on several spatiotemporal physical systems using a common rollout interface. Evaluation separates short-horizon prediction error from long-horizon stability and physical behavior.
Initial experiments establish the end-to-end feasibility of using ShinkaEvolve to generate, execute, and iteratively improve numerical simulation programs under a limited evaluation budget. The evolved artifacts remain explicit source code, making them directly inspectable and potentially easier to adapt than opaque learned surrogates. At the same time, the benchmark exposes practical challenges for evolutionary simulation development, including computationally expensive fitness evaluation, unstable rollouts, and the need for objectives that reward meaningful temporal evolution rather than persistence.
These results position evolutionary code search as a promising tool for constructing and updating numerical components of digital twins, including future ocean and Earth-system applications where models must evolve alongside observations and operational requirements.
Comment
To browse or post comments, you must log in.Log in