NUCLEUS -MoE: Unified Model of Pool Boiling for Liquid Cooling Conference: Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2; August 09–13, 2026; Jeju Island, Republic of Korea Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ’26), August 09–13, 2026, Jeju Island, Republic of Korea DOI: 10.1145/3770855.3819026 ISBN: 979-8-4007-2259-2/2026/08 CCS: Computing methodologies Neural networks CCS: Applied computing Physics CCS: Computing methodologies Model development and analysis

Arthur Feeney, Xianwei Zou, Sheikh Md Shakeel Hassan, Siddhartha Rachabathuni, Aparna Chandramowlishwaran · 2026

Read NUCLEUS -MoE: Unified Model of Pool Boiling for Liquid Cooling Conference: Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2; August 09–13, 2026; Jeju Island, Republic of Korea Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ’26), August 09–13, 2026, Jeju Island, Republic of Korea DOI: 10.1145/3770855.3819026 ISBN: 979-8-4007-2259-2/2026/08 CCS: Computing methodologies Neural networks CCS: Applied computing Physics CCS: Computing methodologies Model development and analysis by Arthur Feeney, Xianwei Zou, Sheikh Md Shakeel Hassan, Siddhartha Rachabathuni, Aparna Chandramowlishwaran (2026) 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-2605-27722.txt.

Introduction

Arthur Feeney, Xianwei Zou, Sheikh Md Shakeel Hassan, Siddhartha Rachabathuni, and Aparna Chandramowlishwaran present NUCLEUS, an expert-mix model for pool boiling. The model jointly handles saturated and subcooled boiling across three fluid groups, replacing sets of specialized surrogates. Expert routing shows coherent spatial specialization without explicit supervision, and the model carries zero-shot and few-shot to new fluids.

Catalog record

CREATORArthur Feeney, Xianwei Zou, Sheikh Md Shakeel Hassan, Siddhartha Rachabathuni, Aparna Chandramowlishwaran
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46689
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.27722

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

NUCLEUS -MoE: Unified Model of Pool Boiling for Liquid Cooling Conference: Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2; August 09–13, 2026; Jeju Island, Republic of Korea Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ’26), August 09–13, 2026, Jeju Island, Republic of Korea DOI: 10.1145/3770855.3819026 ISBN: 979-8-4007-2259-2/2026/08 CCS: Computing methodologies Neural networks CCS: Applied computing Physics CCS: Computing methodologies Model development and analysis Arthur Feeney, Xianwei Zou, Sheikh Md Shakeel Hassan, Siddhartha Rachabathuni, Aparna Chandramowlishwaran Abstract . Two-phase boiling enables heat transfer rates an order of magnitude higher than single-phase cooling, but it remains difficult to model due to the strong coupling between phase change, turbulence, and transport, as well as extreme sensitivity to fluid properties and thermodynamic conditions. Existing learning-based surrogates are either condition- or fluid-specific, limiting generalization and requiring separate models. We present NUCLEUS , a mixture-of-experts model for pool boiling that replaces collections of specialized surrogates with a single architecture. NUCLEUS combines neighborhood attention, signed distance field reinitialization for interface consistency, and expert routing that exhibits emergent specialization across distinct boiling dynamics. Trained on high-fidelity simulations of pool boiling, NUCLEUS jointly models saturated and subcooled boiling across three fluid classes (dielectrics, refrigerants, and cryogens), resolving failure modes of prior models on extreme fluids. We show that expert routing exhibits coherent spatial structure and specialization without explicit supervision. Quantitatively, NUCLEUS matches or exceeds baselines while maintaining physical consistency across heterogeneous boiling configurations. We also show zero-shot and few-shot generalization capabilities on downstream tasks such as a new fluid (Opteon 2P50 developed for immersion cooling). These results demonstrate that mixture-of-experts models are a scalable pathway toward unified surrogate modeling of boiling dynamics and lay the groundwork for broader generalization across scientific ML. Figure 1 . NUCLEUS unifies saturated and subcooled boiling across multiple fluids within a single architecture. (a) Physical fields reveal different dynamics: saturated boiling (top) shows concentrated evaporation and rising large bubbles, while subcooled boiling (bottom) exhibits bulk condensation resulting in turbulent vortices and smaller bubbles. (b) Mixture-of-experts (MoE) routing patterns show emergent specialization suggesting learned decomposition of these distinct transport phenomena. † † cc-license: by 1. Introduction Boiling is the most efficient known form of heat transfer, in which vapor formation at a heated surface removes thermal energy as bubbles rise and detach. This two-phase cooling phenomenon underpins emerging technologies ranging from data-center thermal management to nuclear energy systems ( 9 ) . Growing GPU power densities and rising energy costs have intensified interest in high-performance liquid cooling solutions ( 1 ) . However, boiling dynamics arise from tightly coupled multiphase transport, interface motion, and heat transfer, resulting in highly nonlinear behavior. As bubbles detach and rise from the heater surface, they generate complex vortex structures and turbulent fluctuations in their wakes. The resulting velocity field exhibits high-frequency temporal variations and fine-scale spatial structures that are difficult to model. The design space is also vast: differences in working fluid, bulk temperature, gravity, heater surface, or geometry can fundamentally alter bubble dynamics, flow patterns, and ultimately cooling efficiency. Experimental measurements provide only partial observability of these processes ( 4 ) , while high-fidelity simulations remain computationally expensive ( 7 ; 36 ; 8 ) , motivating data-driven surrogates. Recent learning-based approaches have shown promise for modeling boiling dynamics from simulation data ( 16 ; 20 ) . Yet, existing models are typically trained separately for individual conditions or fluids. Attempts to train unified models jointly across heterogeneous boiling problems often exhibit leakage (e.g., hallucinated condensation "wisps" in saturated boiling) and fail under large variations in thermophysical properties, particularly for extreme fluids such as liquid nitrogen (LN2) ( 17 ) . As a result, current methods produce collections of specialized models and lack a unified representation of boiling physics. In this work, we aim to learn a single model for boiling dynamics that encodes shared structure while specializing across fundamentally different physical behaviors. Generalization in scientific ML requires architectural awareness of when physics fundamentally changes. We show that mixture-of-experts (MoE) architectures naturally align with this setting, exhibiting emergent separation of fluid- and condition-dependent dynamics while preserving shared representations. Combined with physics-informed inductive biases, this results in a single model capable of robust generalization across boiling conditions, where popular conditioning methods (e.g., FiLM ( 31 ) ) fail. Our main contributions are: 1. Unified boiling surrogate. We present NUCLEUS 1 1 1 NUCLEUS code is available at https://github.com/therml-ai/NUCLEUS , a single mixture-of-experts model that jointly learns saturated and subcooled pool boiling across multiple fluids with widely varying thermophysical properties. 2. Physics-informed architectural design. We combine neighborhood attention for local transport, signed distance field reinitialization for interface correctness, and MoE routing to enable specialization across phase-change and flow physics. 3. Interpretability of expert specialization. We show expert routing exhibits coherent spatial structure and emergent specialization to physical phenomena without supervision, providing insight into how boiling dynamics are decomposed within the model. 4. State-of-the-art accuracy with a unified backbone. NUCLEUS matches or exceeds baselines while preserving physical consistency over long rollouts in a unified model, demonstrating that MoE-based surrogates can achieve competitive performance. 5. Generalization to downstream tasks. We demonstrate transfer to a new fluid (Opteon 2P50) unseen during pretraining, highlighting the model’s ability to act as a foundation for multiphase multiphysics modeling with minimal finetuning. 2. Background 2.1. Problem Formulation Boiling is a nonlinear multiphase process governed by the incompressible Navier-Stokes equations coupled with an energy transport equation, solved in both liquid and vapor phases on a spatial domain \Omega\subset\mathbb{R}^{2} evolving over time \mathbb{T}=[0,\infty) . The liquid and vapor phases \Omega_{L},\Omega_{V}\subseteq\Omega are separated by a dynamically evolving interface \Gamma where phase change occurs. We use high-fidelity Flash-X boiling simulations ( 11 ; 8 ) to solve these equations and generate the BubbleML datasets ( 16 ; 17 ) . We also generate new data modeling the dielectric fluid Opteon 2P50. At each timestep, the system is represented by three spatially-resolved fields (Figure 1 ): (i) temperature T(x,t):\Omega\times\mathbb{T}\rightarrow\mathbb{R} , (ii) velocity U(x,t):\Omega\times\mathbb{T}\rightarrow\mathbb{R}^{2} , and (iii) signed distance function \phi(x,t):\Omega\times\mathbb{T}\rightarrow\mathbb{R} , which tracks the phase boundary ( \phi<0 in \Omega_{L} , \phi>0 in \Omega_{V} , and \phi=0 at \Gamma ). Together, these define the state S=(T,U,\phi) . In addition to the fields, the dynamics depend on fluid-specific thermophysical properties and operating conditions, collectively denoted p . These include non-dimensional numbers (Reynolds Re, Prandtl Pr, Stefan St), thermal conductivity, gravity, and bulk liquid temperature. A full list is in Appendix A . These parameters vary significantly across working fluids. Our learning task is to approximate the forward evolution of the state with a timestep of \Delta t : S(\cdot,t+\Delta t):=M(S(\cdot,t),p) . The model M is applied autoregressively to generate longer trajectories: S(\cdot,t+\Delta tn):=M^{n}(S(\cdot,t)) . This problem is challenging because boiling combines stochastic nucleation, turbulent transport, moving interfaces, and localized phase change, resulting in chaotic dynamics and strong spatiotemporal heterogeneity. Long trajectories are inherently sensitive to small perturbations, making pointwise prediction ill-posed. Instead, physically meaningful evaluation relies on statistical field properties and distributions ( 33 ) . 2.2. Physics Heterogeneity and ML Challenges Subcooled vs. Saturated Boiling. Boiling is categorized by whether the bulk liquid is at (saturated) or below

Read the complete text