Bubbleformer: Forecasting Boiling with Transformers
Sheikh Md Shakeel Hassan, Xianwei Zou, Akash Dhruv, Viswanath Ganesan Aparna Chandramowlishwaran · 2025
Read Bubbleformer: Forecasting Boiling with Transformers by Sheikh Md Shakeel Hassan, Xianwei Zou, Akash Dhruv, Viswanath Ganesan Aparna Chandramowlishwaran (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-21244.txt.
Introduction
Sheikh Hassan and co-authors introduce Bubbleformer in 2025, a model that forecasts boiling. Boiling is chaotic and mixes gas and liquid, so it is hard to predict. Past neural models needed future facts, such as bubble spots, to work at all. They also failed on flow boiling speed fields.
The new model uses a transformer design with split attention along each axis. It learns bubble birth, surface change, and heat flow from past states alone. Training spans many fluids, shapes, and settings. The team also release a rich new dataset and fresh physics-based scores for heat flux, surface shape, and mass checks.
Catalog record
| CREATOR | Sheikh Md Shakeel Hassan, Xianwei Zou, Akash Dhruv, Viswanath Ganesan Aparna Chandramowlishwaran |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 87 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-21244
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Bubbleformer: Forecasting Boiling with Transformers Sheikh Md Shakeel Hassan, Xianwei Zou, Akash Dhruv, Viswanath Ganesan Aparna Chandramowlishwaran Abstract Modeling boiling—an inherently chaotic, multiphase process central to energy and thermal systems—remains a significant challenge for neural PDE surrogates. Existing models require future input (e.g., bubble positions) during inference because they fail to learn nucleation from past states, limiting their ability to autonomously forecast boiling dynamics. They also fail to model flow boiling velocity fields, where sharp interface–momentum coupling demands long-range and directional inductive biases. We introduce Bubbleformer , a transformer-based spatiotemporal model that forecasts stable and long-range boiling dynamics including nucleation, interface evolution, and heat transfer without dependence on simulation data during inference. Bubbleformer integrates factorized axial attention, frequency-aware scaling, and conditions on thermophysical parameters to generalize across fluids, geometries, and operating conditions. To evaluate physical fidelity in chaotic systems, we propose interpretable physics-based metrics that evaluate heat flux consistency, interface geometry, and mass conservation. We also release BubbleML 2.0 , a high-fidelity dataset that spans diverse working fluids (cryogens, refrigerants, dielectrics), boiling configurations (pool and flow boiling), flow regimes (bubbly, slug, annular), and boundary conditions. Bubbleformer sets new benchmark results in both prediction and forecasting of two-phase boiling flows. 1 Introduction Boiling is one of the most efficient modes of heat transfer due to the large latent heat of vaporization at liquid-vapor interfaces, making it an attractive solution for ultra-high heat flux applications such as nuclear reactors and next-generation computing infrastructure. Companies such as ZutaCore 1 1 1 https://zutacore.com and LiquidStack 2 2 2 https://www.liquidstack.com are pioneering two-phase and immersion cooling technologies for data centers supporting AI workloads. These industrial efforts reflect a broader trend toward harnessing phase-change phenomena for thermal control in compact, high-density environments. Boiling also holds promise for spacecraft thermal control, but without buoyancy to aid vapor removal, boiling in microgravity faces severe challenges, limiting its current viability. Addressing these limitations will require novel boiling architectures optimized for low gravity, guided by modeling, design optimization, and experimental validation–an iterative process that is computationally intensive and costly. More fundamentally, accurately modeling two-phase pool and flow boiling remains one of the grand challenges in fluid dynamics. The underlying physics is inherently chaotic and multiscale: bubbles nucleate stochastically on heated surfaces, liquid-vapor interfaces continuously deform, coalesce, and break apart, and transitions between flow regimes (e.g., bubbly, slug, annular) occur unpredictably under strong thermal-hydrodynamic coupling. As illustrated in Figure 1 , boiling systems can span a vast range of spatial and temporal scales and are highly sensitive to boundary conditions, fluid properties, and geometry. Figure 1 : (a) BubbleML 2.0 dataset : Visualizes temperature and phase fields across pool and flow boiling configurations. Bubbles nucleate on heated surfaces, deform, coalesce, and transition between regimes, with dynamics strongly coupled to boundary conditions, working fluid, and geometry. (b) Bubbleformer downstream tasks : Autonomous forecasting of full-field dynamics including nucleation and interface evolution, and prediction conditioned on future bubble interfaces. Recent advances in high-fidelity multiphysics solvers such as Flash-X DUBEY2022 ; DHRUV2024113122 have enabled simulations of boiling by solving incompressible Navier-Stokes equations in two phases with interface tracking and nucleation models. Although physically accurate, these simulations are computationally expensive, often requiring days on petascale supercomputers to simulate seconds of physical time DHRUV2019103099 ; DHRUV2021121826 . This has motivated the development of machine learning (ML) surrogates that learn to approximate spatiotemporal evolution of boiling directly from data. These models promise orders-of-magnitude acceleration, enabling new capabilities in real-time forecasting, parametric studies, and design-space exploration. However, current ML models for boiling hassan2023bubbleml ; khodakarami2025mitigatingspectralbiasneural exhibit three main limitations: (1) They require future bubble positions as input to predict velocity and temperature fields, making them unsuitable for forecasting, (2) They fail to learn bubble nucleation, a stochastic discontinuous process central to long rollouts, and (3) They fail to predict velocity fields in flow boiling, even when provided with future bubble positions. We introduce Bubbleformer , a transformer-based spatiotemporal model that forecasts full-field boiling dynamics that includes temperature, velocity, and signed distance fields representing interfaces, setting a new benchmark for ML-based boiling physics. Bubbleformer makes the following core contributions: • Beyond prediction to forecasting. By operating directly on full 5D spatiotemporal tensors and preserving temporal dependencies, Bubbleformer learns nucleation, key to forecasting and predicting long-range dynamics. Unlike prior models that compress time or require future bubble positions, our approach infers them end-to-end. • Generalizing across fluids and flow regimes. We introduce BubbleML 2.0 , the most comprehensive boiling dataset to date, comprising over 160 high-fidelity simulations across diverse fluids (e.g., cryogenics, refrigerants, dielectrics), boiling configurations (pool and flow), heater geometries (single- or double-sided heating), and flow regimes (bubbly, slug, and annular until dryout). Bubbleformer is conditioned on thermophysical parameters, allowing a single model to generalize across these axes. • Physics-based evaluation. We introduce new interpretable metrics to assess physical fidelity beyond pixel-wise error. These include heat flux divergence, Eikonal equation loss for signed distance functions, and conservation of vapor mass. Together, these metrics provide a more rigorous evaluation of physical correctness in chaotic boiling systems. To our knowledge, Bubbleformer is the first model to demonstrate autonomous, physically plausible forecasting of boiling dynamics . It sets new state-of-the-art benchmarks on both prediction and forecasting tasks in BubbleML 2.0, representing a significant step toward practical, generalizable ML surrogates for multiphase thermal transport. 2 Problem Statement Boiling involves the phase change of a liquid into vapor at a heated surface, forming bubbles that enhance turbulence and heat transfer. This phenomenon is highly chaotic: bubbles nucleate unpredictably on the heated surface, then grow, merge, and eventually detach, all while interacting with the surrounding liquid. The boiling system is governed by the incompressible Navier-Stokes equations (for momentum conservation) coupled with an energy transport equation, solved in both the liquid and vapor phases. These equations describe the evolution of the fluid’s velocity \vec{u} and pressure fields P along with the temperature field T that captures heat distribution. The liquid-gas interface is represented by a level-set function \phi (signed distance field) that tracks the moving phase boundary. The interface \Gamma is defined by \phi=0 , with \phi>0 in the vapor region and \phi<0 in the liquid. The governing equations are non-dimensionalized using characteristic scales defined in the liquid phase (e.g., capillary length for length scale and the terminal velocity for velocity scale) and key dimensionless numbers such as Reynolds, Prandtl, Froude, Peclet, and Stefan. To ensure consistency across both phases, vapor properties (e.g., density, viscosity, thermal conductivity, and specific heat) are specified relative to the liquid’s properties. Interfacial physics is modeled by enforcing the conservation of mass and energy at \Gamma , accounting for surface tension and latent heat. Jump conditions for velocity, pressure, and temperature across the interface are implemented using the Ghost Fluid Method (GFM), a numerical scheme that handles sharp discontinuities at the boundary without smearing them. The level-set interface \phi is updated via an advection (convection) equation, which moves the interface with the local fluid velocity. Evaporation and condensation are governed by differences in interfacial heat flux between the two phases, coupling flow dynamics with thermal effects. The two-phase numerical simulation framework is implemented in Flash-X DUBEY2022 and follows the formulation in DHRUV2024113122 . For completeness, the governing equations and numerical modeling