CTF_Panda Model

This directory contains an implementation of the Panda model (Physics-Aware Neural Dynamics) for testing on the CTF for Science Framework, based on the paper "Physics-Aware Neural Dynamics for Learning and Control" (https://arxiv.org/abs/2505.13755).

Files

• ctf_panda.py: Contains the Panda model implementation adapted for the CTF framework.
• run.py: Batch runner script for running the model across multiple sub-datasets.
• config_XX.yaml: Example configuration file for running the model.

Usage

To run the Panda model, use the run.py script from the project root followed by the path to a configuration file. For example:

python models/ctf_panda/run.py models/ctf_panda/config/config_XX.yaml

Model Description

Panda is a physics-aware neural dynamics model that combines machine learning with physical priors to learn dynamical systems. This implementation uses the MLM-trained model from HuggingFace (https://huggingface.co/GilpinLab/panda_mlm).

Configuration Structure

Each configuration file must include the following:

• dataset (required):
  • name: The dataset name (e.g., ODE_Lorenz, PDE_KS).
  • pair_id: Specifies sub-datasets to run on. Formats:
    • Single integer: pair_id: 3
    • List: pair_id: [1, 2, 3, 4, 5, 6]
    • Range string: pair_id: '1-6'
    • Omitted or 'all': Runs on all sub-datasets.
• model:
  • name: Panda
  • zero_shot: Boolean flag to enable/disable fine-tuning (note currently only zero_shot=True is implemented)
  • weights: Path to the pre-trained MLM model (default: HuggingFace GilpinLab/panda_mlm)
  • normalize: Boolean flag to normalise the data with the standard routine provided by Panda's repo
  • context_length: Maximum length of the context for the foundation model, note that context_length=-1 corresponds to using the full warm-start data as context, but this needs to be limited in the PDE case due to computational constraints.

Example configuration:

dataset:
  name: ODE_Lorenz
  pair_id: 1-9 
model:
  name: Panda
  zero_shot: True
  weights: "GilpinLab/panda_mlm"
  normalize: False
  context_length: -1

Requirements

PAnda for CTF relies on Python>=3.10 and the packages listed in requirements.txt.