Research map / Probabilistic and scientific modeling / Dynamical systems
Bubbleformer: Forecasting Boiling with Transformers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Bubbleformer, a factorized space-time axial-attention transformer with FiLM fluid conditioning and frequency scaling that autonomously forecasts boiling dynamics (nucleation included), plus the BubbleML 2.0 dataset and physics-based metrics.
Problem
Neural PDE surrogates for boiling need future bubble positions at inference, fail to learn stochastic nucleation in autoregressive rollouts, and fail to predict velocity fields in flow boiling.
Approach
A spatiotemporal transformer operating on full 5D tensors with factorized space-time axial attention, hierarchical 2x2 patch embedding, FiLM conditioning on a 9D thermophysical descriptor, and frequency-aware attention/feature scaling, trained to jointly predict interface, temperature and velocity fields without simulation input at inference.
Key results
- Bubbleformer outperforms UNet_mod and FFNO on all reported BubbleML metrics in subcooled pool boiling prediction, and stays stable up to 2000 rollout steps.
- On flow boiling forecasting, Bubbleformer-S rolls out 200 steps on an unseen FC-72 trajectory while keeping heat flux distributions, signed distance fields and vapor mass close to simulation; baselines UNet and FNO fail to converge on flow boiling velocity.
- Eikonal losses below 0.1 are reported as sufficient for stable interface evolution under autoregressive rollout.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Bubbleformer instead of UNet_mod and FFNO | subcooled pool boiling prediction · BubbleML subcooled pool boiling | relative L2 error | Helped |
| use Bubbleformer instead of FFNO and UNet_mod for autoregressive prediction | subcooled pool boiling · BubbleML subcooled pool boiling, 800-step rollout | prediction stability | Helped |
| use Bubbleformer for autonomous forecasting | boiling dynamics · unseen flow boiling trajectory, 200-step rollout | physical fidelity | Helped |
| attempt to forecast nucleation with UNet | UNet · boiling autoregressive rollout | nucleation capability | Hurt |
| use FNO for autoregressive forecasting | FNO · boiling autoregressive rollout | rollout stability | Hurt |
| use UNet or FNO to predict velocity in flow boiling | UNet and FNO · BubbleML flow boiling datasets | velocity prediction convergence | Hurt |
Surprising findings
- UNet models perform well in single-step prediction but fail to nucleate bubbles in autoregressive rollouts, while FNO often diverges after a single step.
- FNO and UNet fail to converge when predicting velocity across all flow boiling datasets in BubbleML, even with future interface fields given.
Limitations
- Cannot natively operate on AMR grids; interpolation to uniform grids can introduce numerical errors that destabilize training.
- Specialized to one type of physics; combining datasets (e.g., subcooled and saturated pool boiling) remains challenging.
- High-resolution training demands large GPU memory due to activation storage.
Experimental setup
Datasets: BubbleML 1.0 BubbleML 2.0
Benchmarks: BubbleML 2.0 subcooled pool boiling prediction BubbleML 2.0 flow boiling forecasting
Metrics: RMSE relative L2 error BRMSE IRMSE Fourier mode errors heat flux KL divergence Eikonal loss relative vapor volume error
Baselines: UNet FNO UNet_mod FFNO
Builds on: BubbleML Factorized Fourier Neural Operator UNet FiLM T5 relative position embeddings
Closest papers by meaning
- Neptuna: A Comprehensive Machine Learning Framework for Benchmarking Complex Multiphase Flows (NeurIPS 2026 · similarity 0.85)
- Interactive 4D Volumetric Liquid Forecasting under Moving-Solid Interaction (NeurIPS 2026 · similarity 0.84)
- Calibrated Physics-Informed Uncertainty Quantification (ICML 2025 · similarity 0.82)
- PINP: Physics-Informed Neural Predictor with latent estimation of fluid flows (ICLR 2025 · similarity 0.82)
- EddyFormer: Accelerated Neural Simulations of Three-Dimensional Turbulence at Scale (NeurIPS 2025 · similarity 0.82)
- CALM-PDE: Continuous and Adaptive Convolutions for Latent Space Modeling of Time-dependent PDEs (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.