Trained models for "Missing nucleation in learned pool-boiling forecasters: a squared-error cause and a phase-indicator remedy"
These are the 24 forecasters trained for the paper. Each is a Bubbleformer model (axial vision transformer, FiLM-conditioned transformer or modern U-Net). The models were trained on saturated pool boiling (FC-72) or on the single-bubble control, both from BubbleML 2.0 (DOI 10.57967/hf/5594).
run<id>.ckpt: the final checkpoint (epoch 150). It holds the weights (state_dict) and the training configuration (hyper_parameters); optimiser states are removed.manifest.csv: run id, the paper's label, architecture, seed, file size and SHA-256 for each file.
The run registry (configs/runs.csv), evaluation scripts and reproduction guide are in the code
repository: https://github.com/TheNameisSea/pool-boiling-nucleation. Evaluate a model with:
bin/evaluate_run.sh <run> run<run>.ckpt
The weights are licensed CC BY-NC-SA 4.0 (see LICENSE), the licence of the training data, BubbleML 2.0
(revision 262b6556 on Hugging Face). The code is MIT.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support