How Machine Learning Predicts Fluid Densities under Nanoconfinement

Yuanhao Li · 2025

Read How Machine Learning Predicts Fluid Densities under Nanoconfinement by Yuanhao Li (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-2508-17732.txt.

Introduction

A fluid squeezed into a pore a few molecules wide layers itself against the walls, and that layering governs flow and chemistry in tight spaces. Yuanhao Li shows that a machine learning model can learn the pattern from simulation data and then predict it cheaply. A random forest trained on large molecular simulation sets reproduced the uneven density profiles across the pore.

The model interpolated excellently across temperatures and pore widths it had seen, and even extrapolated modestly beyond them. Decades of theory have failed to give a general formula for this layering, leaving expensive simulations as the default. Li offers the learned surrogate as a lower-cost, broadly usable stand-in for practitioners who model fluids in nanopores, from energy storage to filtration.

Catalog record

CREATORYuanhao Li
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-17732

Browse the full library · About the Tower of Babel

Full text

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

How Machine Learning Predicts Fluid Densities under Nanoconfinement Yuanhao Li Abstract Fluids under nanoscale confinement differ – and often dramatically – from their bulk counterparts. A notorious feature of nanoconfined fluids is their inhomogeneous density profile along the confining dimension, which plays a key role in many fluid structural and transport phenomena in nanopores. Nearly five decades of theoretical efforts on predicting this phenomenon (fluid layering) have shown that its complexity resists purely analytical treatments; as a consequence, nearly all current approaches make extensive use of molecular simulations, and tend not to have generalizable predictive capabilities. In this work, we demonstrate that machine-learning-based models (in particular, a random forest model), trained upon large molecular simulation data sets, can serve as reliable surrogates in lieu of further molecular simulation. We show that this random forest model has excellent interpolative capabilities over a wide range of temperatures and confining lengthscales, and even has modest extrapolative ability. These results provide a promising pathway forward for developing models of nanoconfined fluid properties that are generalizable, lower cost than “pure” molecular simulation, and sufficiently predictive for fluids-in-nanopores practitioners. I Introduction As compared to their bulk counterparts, fluids under nanoscale confinement are imbued with a host of remarkable properties NanofluidicsComingOfAge ; ChengzhenNCF2020 – including anomalous diffusion DiffusivityVsDensity ; LayeringDependentDiffusion ; ABF-NRA-DiffusionCNTs ; FirstLayerDiffusivity ; li2022produce , viscosity NeekAmalViscosity , thermal phenomena FLGKapitza ; PoulikakosWaterThermalConductivity ; NagoeCp ; kim2008thermal , pressure distribution ShiPressureCNT ; ZouMarooPressure , phase behavior Koga2000 ; TanakaBilayerIcePhase ; TanakaWaterCNTPhaseDiagram ; ABF-NRA-MultipleWaterPhases ; AgrawalStranoWaterCNT , contact-line dynamics LichterMovingContactLine , and slip SendnerBocquet ; BocquetInterlayerExchange ; BarratBocquet_LargeSlipEffect ; BocquetMicroNanoFlowReview ; LichterFK ; LichterFK_JFM ; LichterSlipRateProcess , amongst many others – which all fundamentally stem from the comparability of an internal (molecular) lengthscale to the confinement lengthscale. At the heart of many of these anomalous properties is the inhomogeneous fluid density profile that emerges under confinement. Unlike bulk (single-phase) fluids, which are constant in density throughout space, nanoconfined fluids often feature significant density anisotropy, namely, long-ranged correlations along the direction of confinement. These correlations manifest as ordered layers of fluid that run parallel to the confining boundary. There is a rich history of studying fluid layering phenomena at interfaces – spanning nearly five decades – using simulation ParsonageLayering ; AbrahamLayering ; ToxvaerdLayeringI ; ToxvaerdLayeringII , theory ToxvaerdLayeringII ; OZInhomogeneous ; RosenfeldFMT ; StillingerBuffInhomogeneousFluids ; ClusterExpansionInhomogeneousFluid ; sisan2011end ; AluruEQT ; FluidsInCNTs ; FirstLayerDensity , and experiments YangZewail ; Uhlig2019 . These studies have conclusively established the existence of fluid layering in a broad range of nanopore systems, characterized the lengthscales and density magnitudes associated with layering, and established connections between layering and other fluid properties under nanoconfinement. Despite this progress, one simple challenge continues to loom large: Given a specific choice of fluid and solid, and a set of thermodynamic conditions, can one produce an accurate estimate for the layered density profile, without resorting to time-intensive molecular simulations (or even more resource-intensive experimental characterization techniques)? On this question, we make two key observations: 1. It is enormously difficult to develop a physically motivated, broadly generalizable, and predictive model for fluid layering. There are numerous approaches, grounded in rigorous theory, that (in theory) could be used to compute these density profiles; however, in practice, none of these approaches can generate a meaningful prediction using only simple details about the system setup. In particular, every existing theory contains at least one “rug parameter” (so termed because a large amount of complex physics is “swept under” this parameter). Such parameters are challenging (or impossible) to compute ab initio ; in fact, it is often the case that rug parameters cannot even be coarsely estimated without extensive amounts of simulation (followed by statistical fitting). For example, integral-equation methods (most notably, Ornstein-Zernike Hansen ) can in principle be used to calculate fluid density profiles in the presence of the external potentials induced by the confining boundaries OZInhomogeneous ; however, these methods require detailed a priori knowledge of the homogeneous radial distribution function (not to mention a suitable closure relation). Approaches based upon (liquid-state) Density-Functional Theory, such as Fundamental Measure Theory RosenfeldFMT , require ever-increasing numbers of non-obvious terms in the density functional (along with non-obvious weightings) in order to reproduce density profiles in systems of modest complexity. The Achilles heel of cluster-expansion techniques StillingerBuffInhomogeneousFluids ; ClusterExpansionInhomogeneousFluid is their reliance on virial coefficients. More recent “semi-classical” theories AluruEQT ; FirstLayerDensity require detailed information about system-averaged energetics for fluid-solid and fluid-fluid interactions; this is occasionally feasible in simple geometries (employing hard-fought integrals), but not in general an analytically tractable demand. This limitation was clear even in the first-ever thermodynamic treatment of the layering phenomenon ToxvaerdLayeringII , in which the Gibbs surface excess energy played the role of the rug. 2. There are significant practical benefits to simply knowing the inhomogeneous density profile. For example, knowledge of the layered fluid structure is valuable for the spatial calibration of two-viscosity models MyersTwoViscositySlip ; ThomasMcGaugheyFastWaterTransport ; MH-NRA-EnhancementFactor , which describe a host of nanoconfined fluid transport phenomena as originating from two “distinct” fluid phases adjoining each other within the system. Models for anomalous thermal properties of nanoconfined fluids (e.g. Kapitza resistance FLGKapitza or heat capacity NagoeCp ) are often based upon knowledge of the vibrational modes supported by the interfacial fluid layer, which in turn requires knowledge of the density and spatial extent of this layer. Approximate knowledge of the equilibrium nanoconfined fluid density profile can also accelerate molecular simulations, by providing initial conditions that require less evolution in order to attain a state of statistical equilibrium. In other words, even in the absence of any first-principles rationalization, there are numerous applications that would benefit from rapid and accurate estimates for nanoconfined fluid density profiles. Based upon these observations, it is clear that there is a significant need for fast, predictive tools that describe nanoconfined fluid structure; this need is unmet (and unlikely to be met) by existing methods. The premise for the present work naturally follows: A machine-learned (ML) surrogate model for fluid layering, trained using an extensive data set from molecular simulations of confined fluids under a wide variety of degrees of confinement and thermodynamic conditions, has tremendous practical value. We have developed several such ML models and, as we describe below, we find that these models, once trained, are capable of producing accurate predictions for fluid-layering phenomena (and at negligible computational cost as compared to the underlying molecular simulations). Moreover, although our models (like many ML models) are not fundamentally grounded in physical principles, we demonstrate that these models respect relevant physical constraints. Unlike the existing approaches detailed above, we are optimistic that ML-driven models have a clear pathway forward for solving the core challenge of predicting fluid density profiles in a broad range of nanoconfined systems. II Methods In the discussion that follows, we focus our efforts on modeling the density profiles of water confined within graphene nano-slits. Predicting these density profiles is a notoriously challenging problem, especially in the narrowest slits that accommodate water imbibition (0.4 nm); this is both because the number of fluid layers in the channel can change suddenly as the channel increases in width, and also because the layered structures induced by each confining boundary can significantly

Read the complete text