Direct numerical simulation of turbulent heat transfer in water with temperature-dependent viscosity and conductivity

arXiv:2611.14276 · physics.flu-dyn · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Direct numerical simulation of turbulent heat transfer in water with temperature-dependent viscosity and conductivity".

Title and authors: Kai: Now let's look at what they actually found in this paper about turbulent heat transfer in water with temperature-dependent viscosity and conductivity. They used a direct numerical simulation to see how the heat moves when the fluid properties aren't constant.

Mira: The main finding is that the standard picture of a passive scalar being transported with just one constant Prandtl number doesn't work anymore because those property variations are too big.

Lev: So what does that mean practically? Does it just mean we have to throw in more empirical corrections for every single simulation run, which sounds computationally expensive?

Kai: Actually, they developed an integral transformation framework. This lets them map the variable-property profiles onto constant-property ones by using the mean momentum and thermal balances to derive those transformations.

Mira: And the result of these inverse transformations is that they predict friction and heat transfer within two percent of the direct numerical simulation results, except at very low Reynolds numbers, which is where things get messy.

Lev: Two percent error across such a wide range of conditions sounds quite good for a predictive tool, but what's the catch? The paper mentions limitations or caveats that we need to keep in mind when we think about implementing this on actual hardware.

Kai: They do point out that the classical picture of just using one Prandtl number breaks down completely when you have these property variations, so it’s a big conceptual shift.

Mira: And they also mention how the viscosity of liquid water is especially sensitive, halving between three hundred and three hundred forty Kelvin and varying by a factor of six between freezing and boiling point. That sensitivity is what drives the need for this more complex approach.

Lev: If we were designing an error correction system for a real thermal management chip, knowing that the underlying physics requires mapping variable properties to constant ones sounds like it adds layers of complexity to the calibration process.

The paper's summary: Kai: So summarizing what they did, this paper focuses on solving the Navier-Stokes equations with temperature-dependent viscosity and conductivity for liquid water in a plane channel using direct numerical simulation. They track wall heating and cooling across a wide range of temperatures and Prandtl numbers.

Mira: The core idea is to use integral transformations derived from the mean momentum and thermal balances to create inverse transformations that link the variable property profiles to simpler, constant-property ones.

Lev: So, if we are trying to run this on a real system, what does it actually take? Do we need a massive computational budget just to calculate those transformation kernels before we even start the main simulation?

Kai: The benefit is that these inverse transformations predict friction and heat transfer within two percent of the direct numerical simulation results, provided you stay above the lowest Reynolds numbers. That level of accuracy across that range is what they’re showing.

Mira: They also show how this framework extends to liquids by accounting for the decoupling between conductivity and viscosity and the variation in Prandtl number across the layer, which goes beyond just applying simple empirical corrections to constant-property formulas.

Lev: That extension into liquids is important because as I mentioned, conductivity doesn't scale nicely with viscosity, so a single correction factor isn't going to cut it for real hardware fidelity.

Kai: They also provide some insights into near-wall physics by showing that instantaneous fields and spectra reveal that the wall friction Reynolds number doesn't accurately represent the range of turbulent scales.

Mira: And they found something interesting about the near-wall streaks, where velocity and temperature have a common spacing when scaled with the local viscous length. The correlation between those fluctuations is strong, being plus zero point eighty five for wall cooling and minus zero point eighty five for wall heating.

The paper's improvements: Lev: Okay, so they've done the main simulation and found these transformations. What are the actual suggested improvements or ways to make this framework better? They don't just stop at the DNS results, right?

Kai: They extend the modeling variable properties into Reynolds-averaged Navier-Stokes models like Spalart and Allmaras. Specifically, they carry over eddy viscosity and eddy diffusivity corrections from section five point three of that earlier work by Ceci et al.

Mira: And they also modify the damping function in that model by introducing terms related to nu itself through f sub two, which essentially means you can introduce these property corrections without having to recalibrate everything from scratch.

Lev: That sounds like a practical step for implementation, moving this from pure simulation insight to something we could actually plug into a standard RANS solver for real-world design checks. It reduces the initial overhead.

Kai: They also show that when comparing constant-property flow at the same bulk Reynolds number based on the mean viscosity, wall cooling increases the friction coefficient by up to thirty nine percent when the temperature ratio is one point three to one.

Mira: And wall heating reduces it by up to thirty percent to thirty nine percent at a temperature ratio between zero point eight three and zero point eight five. These are these specific quantitative results that show how much those property variations impact the friction factor in real scenarios.

Lev: Those percentage changes are substantial, especially when you consider that we're talking about corrections across the entire range of conditions with only a few percent error according to their own testing against constant-property flows. That’s a lot of leverage for model validation.

Kai: The authors also note that the property laws themselves introduce new dimensional parameters, like the wall temperature first, so that two flows with the same bulk Reynolds number and Prandtl number can differ just based on how differently they are heated at the wall.

Conclusion: Mira: So to wrap up on this paper, the key takeaway is that for turbulent heat transfer in water with variable properties, you need a predictive model that can handle those property variations near the wall.

Kai: The integral transformations derived from mean momentum and thermal balances are a powerful way to map those variable profiles onto constant ones, giving predictions within two percent of direct numerical simulation results for friction and heat transfer.

Lev: And they managed to extend this framework to liquids by correctly accounting for the decoupling of conductivity from viscosity and the variation in Prandtl number across the layer, which is a necessary step before we can even think about using it on real hardware.

Kai: They showed how these corrections can be integrated into RANS models like Spalart and Allmaras by using heat transfer formulations, making it feasible for use in industry applications.

Mira: The paper also highlights that the effect of all these property variations is mostly confined to a thin layer at the wall, which is exactly what a predictive model needs to describe accurately.

Lev: So, we have a framework that connects universal passive-scalar problems to this more complex one, but they still leave some open questions about extending these transformations to things like pipes and ducts where the outer layer behaves differently than in a channel.

Kai: It’s an interesting piece of work because it gives us concrete tools for handling fluid physics in systems where properties are changing significantly.

Mira: Indeed, the study of "Direct numerical simulation of turbulent heat transfer in water with temperature-dependent viscosity and conductivity" provides a solid foundation for how we model complex fluid behavior.

Davide Modesti, Alessandro Ceci

Gran Sasso Science Institute

physics.flu-dyn

Submitted: 2026-11-10

Updated: 2026-11-10

Importance score: 79/100

The gist: Direct numerical simulation of turbulent heat transfer in water with temperature-dependent viscosity and conductivity clarifies several aspects of the physics The inverse transformations predict…

Key concepts

Fluid Property Sensitivity
Liquid water's viscosity changes drastically between freezing and boiling points. This means the molecular Prandtl number (Pr) can vary by a factor of up to six across a heated or cooled duct, invalidating simple models that assume a single constant property.
Integral Transformation Framework
This mathematical method takes complex profiles with variable properties and transforms them into simpler profiles with constant properties. This allows researchers to use established solutions for simpler flows while accurately predicting friction and heat transfer within 2% of the detailed DNS simulation.
Near-Wall Physics Insights
Analysis of instantaneous fields shows that velocity and temperature streaks near the wall have a consistent spacing when scaled by local viscous length. The correlation between these streaks (+0.85 for cooling, -0.85 for heating) provides crucial information about the physics occurring right at the boundary.
Model Extension to RANS
The property corrections derived from the integral transformations can be incorporated into Reynolds-Averaged Navier-Stokes (RANS) models, such as Spalart & Allmaras. This allows these simpler, computationally cheaper models to account for variable properties without needing extensive recalibration.

Terminology

Summary

Direct numerical simulation of turbulent heat transfer in water with temperature-dependent viscosity and conductivity clarifies several aspects of the physics The inverse transformations predict friction and heat transfer within 2 % of the DNS, except at the lowest Reynolds numbers This approach, developed for air, is extended to liquids, whose conductivity is not proportional to the viscosity and whose Prandtl number varies across the channel

Fluid Property Sensitivity

The viscosity of liquid water is particularly sensitive, as it halves between 300 and 340 K and varies by a factor of six between the freezing and the boiling point The molecular Prandtl number Pr = μc p/λ can therefore change by a factor of up to six between the wall and the core of a heated or cooled duct, and the classical picture of a passive scalar transported at a single Prandtl number does not apply

Modeling Variable Properties

The authors present an integral transformation framework that maps variable-property profiles onto constant-property ones through integral transformations derived from the mean momentum and thermal balances The inverse transformations predict friction and heat transfer within 2 % of the DNS, except at the lowest Reynolds numbers This framework is extended to liquids, accounting for the decoupling of conductivity from viscosity and the variable Prandtl number

Near-Wall Physics Insights

Instantaneous fields and spectra show that the wall friction Reynolds number misrepresents the range of turbulent scales, whereas near-wall streaks of velocity and temperature have a common spacing when scaled with the local viscous length The sign of the correlation coefficient between u' and theta' in the plane is +0.85 for wall cooling and −0. p. 085 for wall heating

Predictive Accuracy

The predictions of the inverse transformations follow the DNS for all property conditions, and the prediction errors are reported in figure 11 and table 2 With respect to constant-property flow at the same Re b (based on μm), wall cooling increases the friction coefficient, by up to 28–32 % at T m/T w = 1.3, while wall heating reduces it, by up to 30–39 % at T m/T w = 0. p. 683–0.85 The corrected model offers a uniform accuracy of a few per cent across the whole range of conditions at no additional cost

Model Extension to RANS

The eddy-viscosity and eddy-diffusivity corrections of § 5.3 can be carried over to RANS models, such as the one-equation model of Spalart & Allmaras (1994) with the heattransfer extension proposed by Ceci et al. p. 2 (2025) The modified damping function also enters S tilde through fv2 and hence nu itself This suggests introducing the property corrections in the same way, without any recalibration

Conclusion on Physical Mechanism

The whole effect of the property variations on friction and heat transfer is thus confined to a thin layer at the wall, which is what a predictive model has to describe The integral transformations derived for air carry over to liquids with minor changes, which account for the decoupling of conductivity from viscosity and for the variable Prandtl number The Spalart–Allmaras model deserves credit in this respect The property laws introduce dimensional parameters, the wall temperature first of all, so that two flows with the same Reynolds number, Prandtl number and temperature ratio differ if their wall temperatures differ This should not discourage its study The underlying equations hold regardless, and the transformations use them to connect the universal passive-scalar problem to this more complex one Several questions remain open, such as the extension of the transformations to pipes and ducts The heating is uniform and symmetric, while heat exchangers are typically heated on one side, with a non-uniform heat flux along the wall The extension of the transformations to pipes and ducts, and to boundary layers, whose outer layer differs from that of the channel, remains to be done The data and the code that generate every figure of the paper are available as JFM Notebooks Each notebook contains the time-averaged DNS statistics, spectra and instantaneous planes shown in the figure, the constant-property reference profiles, the predictions of the transformations and the solutions of the Spalart–Allmaras model The property laws and the inputs of the simulations are given in § 2 The paper is 2611.14276.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Incorporate integral transformations derived from mean momentum and energy balances into AI systems to predict friction and heat transfer coefficients within 2 percent of Direct Numerical Simulation results, except at the lowest Reynolds numbers.

  2. Develop predictive models that map variable-property profiles onto constant-property ones using inverse transformations, achieving predictions within 2 percent of DNS for friction and heat transfer, except at the lowest Reynolds numbers.

  3. Implement corrections derived from integral transformations to predict friction and heat transfer within 1 to 2 percent of the DNS results for both air and liquids, including extensions for liquids where conductivity is not proportional to viscosity.

  4. Integrate empirical corrections, such as the factor (mu m/mu w)n of Sieder & Tate (1936), applied to the Nusselt number, into AI systems for liquid flows with variable properties.

  5. Extend established frameworks developed for air convection with variable properties to liquids by accounting for the decoupling of conductivity from viscosity and the variation in Prandtl number across the layer.

  6. Utilize semi-local scaling, where wall distance is scaled with local viscous length (y+ = yu t/nu w), to collapse mean velocity and shear stresses of liquid-like and gas-like fluids, even though turbulence intensities are not collapsed.

  7. Apply the method of inverse transformations to predict mean velocity and temperature profiles from constant-property counterparts, iteratively updating kernels with temperature through property laws until target bulk Reynolds number and temperature ratio are matched.

  8. Use eddy viscosity and diffusivity models that are functions of wall distance (y+ cp) to account empirically for property variations in RANS models, using coefficients phi and beta derived from DNS data.

  9. Modify the Spalart-Allmaras model using the heat-transfer formulation by introducing eddy viscosity and diffusivity corrections based on integral transformations to improve near-wall asymptotic behavior.

  10. Adjust the turbulent Prandtl number in RANS models by incorporating property corrections into it, where Prt = nu t / alpha t, to account for variable properties in the near-wall region.

Related papers