Deep Learning-based Algebraic Reynolds Stress Closures for RANS Simulations of Turbulent Flows

summary

The gist

Turbulence is ubiquitous in engineering and science, yet direct simulation is prohibitively expensive The Reynolds-averaged Navier-Stokes (RANS) equations provide computational savings exceeding ten…

In short

The Deep Algebraic Reynolds Stress Model (DARSM) is a physics-derived deep learning closure for RANS simulations of turbulent flows. It uses a neural network to model empirical parameters within an implicit algebraic Reynolds stress equation, ensuring physical structure. Trained efficiently using a hybrid discrete adjoint method, DARSM generalizes well across different flow regimes and geometries with high accuracy.

Key concepts

RANS Equations
Reynolds-averaged Navier-Stokes (RANS) equations are used to model turbulent flows by averaging the fluid motion over time. While computationally cheaper than direct simulation, they introduce an unclosed term representing turbulent stresses that requires a closure model to solve.
Algebraic Reynolds Stress Model (ARSM)
This is a specific type of algebraic closure where the relationship between mean flow and Reynolds stresses is defined by an implicit algebraic equation. DARSM uses this structure, but instead of fixing all coefficients from data, it only learns the unknown empirical parameters within this physics-based framework.
Hybrid Discrete Adjoint
This is a specialized method for training deep learning models on complex Partial Differential Equations (PDEs). It combines reusing a fast forward numerical solver with automatic differentiation to calculate gradients efficiently. This allows the model to be trained end-to-end over the entire system, even when standard methods fail due to equation stiffness.
Physics-Derived Closure
DARSM embeds physical knowledge directly into the neural network's structure. Instead of a general network learning everything, it maps flow invariants (physical properties) to empirical parameters within an algebraic equation derived from the Reynolds stress transport equations, imposing strong physical constraints on what the AI learns.

Terminology used across episodes

This episode discusses

The paper

Deep Learning-based Algebraic Reynolds Stress Closures for RANS Simulations of Turbulent Flows · Read on arXiv

Daniel Dehtyriov, Jonathan F. MacArt, Justin Sirignano

University of Oxford · University of Notre Dame

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: "Deep Learning-based Algebraic Reynolds Stress Closures for RANS Simulations of Turbulent Flows".

Mira: Here's the gist of the paper. Turbulence is ubiquitous in engineering and science, yet direct simulation is prohibitively expensive The Reynolds-averaged Navier-Stokes (RANS) equations provide computational savings exceeding ten orders of magnitude but introduce unclosed terms requiring modelling (the closure problem) Offline-trained machinelearning (ML) closures suffer distribution shift when deployed in predictive simulations, while ML methods that bypass the governing equations struggle to generalise from scarce high-fidelity data We develop a physics-derived deep learning closure model for RANS, the Deep Algebraic Reynolds Stress Model (DARSM), which can be trained on small datasets and accurately generalise across Reynolds numbers, to unseen geometries, and to different flow regimes A neural network maps flow invariants to empirical parameters in an implicit algebraic Reynolds stress equation, derived from the Reynolds stress transport equations under the weak-equilibrium assumption, imposing significant physics-based structure on the ML closure DARSM reduces average test velocity error as compared to the baseline RANS by 2–4× across Reynolds number, geometries, and flow regimes, with peak case-level reductions of 12× Model Architecture and Physics Foundation DARSM is a physics-derived deep learning anisotropic closure for RANS that models empirical parameters of an implicit algebraic closure equation derived from the Reynolds stress transport equations (RSTE) under the weak-equilibrium assumption The neural network maps flow invariants to empirical parameters in an implicit algebraic Reynolds stress equation, which fixes the tensorial structure of the closure from physics This approach differs from tensor-basis neural networks (TBNN) by deriving the coefficient structure from physics rather than learning it entirely from data The algebraic Reynolds stress model (ARSM) derives the coefficient structure using physical arguments, making the neural network model only unknown empirical parameters within that physics-based framework Training Strategy and Gradient Derivation To address distribution shift, DARSM is trained by optimizing over the entire PDE system of the RANS momentum equations, continuity equation, and coupled implicit algebraic closure equations Due to the stiffness of the PDE system, standard automatic differentiation (AD) methods numerically fail We derive a system of implicit adjoint equations designed to leverage the forward numerical solver’s implicit-explicit structure to provide computationally efficient optimization for training DARSM This method is called a "hybrid discrete adjoint" because step (1) reuses the forward numerical solver in transpose form, while step (3), which covers the assembly of Hm, gm from the state and parameters, uses AD The hybrid adjoint method yields highly computationally-efficient calculation of the gradients at O(N) memory Generalization and Performance DARSM demonstrates strong generalization capabilities, improving upon the baseline RANS EARSM across all in-sample and out-of-sample flows DARSM trained on one or two DNS cases generalises across an order of magnitude in Reynolds number, across duct aspect ratios, and across periodic-hill geometries, reducing the mean out-of-sample velocity error against the baseline RANS by 2×, 3× and 4× Notably, the duct-trained model transfers without retraining to the periodic-hill case with accuracy on par with a closure trained directly on the periodic hills flow cases DARSM is the only method tested that improves on the baseline both in-distribution (duct) and out-of-distribution (hills) Methodological Advantages The key advantage of DARSM is that it combines a closure whose tensor structure and invariant basis are derived from first principles with a steady-state adjoint that provides solver-consistent gradients at O(N) memory The constant-memory steady-state adjoint formulation requires the forward solver to converge to a steady state, consistent with RANS’s premise that the time-averaged mean of the (unsteady) flow is stationary This construction extends to any semi-implicit PDE solver and allows for highly efficient calculation of gradients at O(N) memory A deep learning closure model, DARSM, is developed for RANS which uses neural networks to model the empirical constants of an algebraic Reynolds stress equation derived under weak-equilibrium The model is trained by optimizing over the entire PDE system using discrete semi-implicit adjoint equations Performance is evaluated on two widely-used benchmarks for turbulence models: the square duct and periodic hills canonical flows DARSM, when trained only on one or two cases, generalises across Reynolds number, duct aspect ratio, periodic-hill geometry and flow regime, including a transfer from attached duct flow to separated hill flow without retraining; it reduces average out-of-sample velocity error as compared to the RANS baseline by 2–4×, with peak case-level reductions of 12× The key is the combination of a closure whose tensor structure and invariant basis are derived from first principles with a steady-state adjoint that provides solver-consistent gradients at O(N) memory Limitations The constant-memory steady-state adjoint formulation requires the forward solver to converge to a steady state, consistent with RANS’s premise that the time-averaged mean of the (unsteady) flow is stationary Genuinely unsteady flows where no time-stationary mean exists are better treated by scale-resolving methods such as LES, for which unsteady adjoint formulations have been developed All results are on 2D grids The weak-equilibrium assumption underlying the EARSM itself has known limits: flows with strong streamline curvature, rapid acceleration, or substantial non-equilibrium dynamics may fall outside its range of validity Numerical Validation The hybrid adjoint method is verified against central finite differences using the vector–Jacobian identity ⟨uˆout, duout⟩ = ⟨uˆin, duin⟩ Each momentum half-sweep (u, v, w in both x and y), the coupled k–ω block sweeps, the pressure projection, the EARSM block, the adaptive forcing, and the composed forward step passes at < 10−5 relative error in double precision The end-to-end gradient check is in Figure 10 Comparison with Existing Methods DARSM outperforms five established ML methods in scientific machine learning for the square duct and periodic hills flow cases DARSM is the only method that improves on the baseline RANS across all in-sample and out-of-sample flows (Section 4) The key is the combination of a closure whose tensor structure and invariant basis are derived from first principles with a steady-state adjoint that provides solver-consistent gradients at O(N) memory A Physics-derived Closure The governing equations as physics-based structure for learning The alternative methods we benchmark each omit one of the structural ingredients of DARSM DARSM is the only method in our benchmark that combines physics-derived closure structure, invariant-aligned coordinates, and adjoint training through the solver A deep learning closure model, DARSM, is developed for RANS which uses neural networks to model the empirical constants of an algebraic Reynolds stress equation derived under weak-equilibrium The model is trained by optimizing over the entire PDE system using discrete semi-implicit adjoint equations Performance is evaluated on two widely-used benchmarks for turbulence models: the square duct and periodic hills canonical flows DARSM, when trained only on one or two cases, generalises across Reynolds number, duct aspect ratio, periodic-hill geometry and flow regime, including a transfer from attached duct flow to separated hill flow without retraining; it reduces average out-of-sample velocity error as compared to the RANS baseline by 2–4×, with peak case-level reductions of 12× The key is the combination of a closure whose tensor structure and invariant basis are derived from first principles with a steady-state adjoint that provides solver-consistent gradients at O(N) memory Acknowledgments and Disclosure of Funding Daniel Dehtyriov’s fellowship is supported by the Schmidt AI in Science Fellowship at the University of Oxford This work is supported by the U.K.

Kai: Engineering and Physical Sciences Research Council grant EP/X031640/1 and the U.S.

Mira: National Science Foundation under Award CBET-22-15472 This research used resources of the Oak Ridge Leadership Computing Facility at the Oak Ridge National Laboratory, which is supported by the Office of Science of the U.S.

Kai: Department of Energy under Contract No.

Title and authors: Kai: So what does this DARSM actually do in practice? The summary says they developed this model that can be trained on relatively small datasets, but the promise is that it can then generalize pretty well across different Reynolds numbers and even to unseen geometries and flow regimes without needing a complete retraining cycle.

Mira: That’s the key takeaway for me. The authors are tackling the distribution shift problem that usually plagues machine learning closures in scientific simulations, where you train on one set of data—say, a square duct—and it fails when you try to predict something like a periodic hill flow

Pope, two thousand: <ref:2605.26358#pg2>.

Lev: They address this by training the neural network not just on the final solution, but by optimizing over the entire system: all those momentum equations, continuity equation, and even those coupled implicit algebraic closure equations

Pope, two thousand: <ref:2605.26358#pg2>.

Kai: And how they do that is through a clever training strategy involving discrete adjoint equations. They use this to get gradients efficiently across the whole system, which is crucial because standard automatic differentiation just breaks down at the grid sizes you actually need for practical turbulence predictions

Pope, two thousand: <ref:2605.26358#pg2>.

Mira: What’s really interesting is how they leverage the structure of the iterative PDE solver itself. They re-use those fast numerical solvers for implicit steps when calculating the adjoint steps in backpropagation, which gives them exact gradients at constant memory, independent of the number of variables

Pope, two thousand: <ref:2605.26358#pg2>.

Lev: From an engineering standpoint, that constant memory gradient calculation is huge. It means you don't need massive amounts of RAM just to train this model effectively on real hardware for solving those RANS problems.

The paper's summary: Kai: So, beyond just creating a closure model, the paper focuses heavily on *how* they built it better than previous attempts. They emphasize that DARSM is physics-derived because the neural network maps flow invariants to empirical parameters within an implicit algebraic equation derived from the Reynolds stress transport equations

Pope, two thousand: <ref:2605.26358#pg2>.

Mira: That’s important because it gives the model a strong inductive bias. It’s not just learning arbitrary relationships; it’s constrained by a set of physical rules derived from the equations themselves, which is what sets it apart from tensor-basis neural networks

Pope, two thousand: <ref:2605.26358#pg2>.

Lev: And that constraint is what allows it to learn mappings instead of just inferring every single piece of physics from very scarce data. It’s like giving the learning algorithm a skeleton based on known laws.

Kai: They compare their results against the baseline RANS EARSM, and they claim that DARSM improves the average test velocity error by two to four times across Reynolds number variations, duct aspect ratios, and even different flow regimes

Pope, two thousand: <ref:2605.26358#pg2>.

Mira: That improvement is quite strong. They also noted peak case-level reductions of twelve times in some specific scenarios

Pope, two thousand: , which suggests that when the model gets it right for a certain case, the physics-informed structure really pays off <ref:2605.26358#pg2>.

Lev: But we have to keep an eye on those conditions where they claim this works. The paper mentions that the weak-equilibrium assumption underlying their algebraic Reynolds stress model has known limits, specifically with flows having strong streamline curvature or rapid acceleration

Pope, two thousand: <ref:2605.26358#pg2>.

The paper's improvements: Kai: So to wrap up the Deep Learning-based Algebraic Reynolds Stress Closures for RANS Simulations of Turbulent Flows paper. The main point is that they’ve created a model that uses deep learning to model the empirical constants of an algebraic Reynolds stress equation derived under weak-equilibrium, and they trained it using discrete semi-implicit adjoint equations.

Mira: And the result is a closure that generalizes across Reynolds number variations, duct aspect ratios, and periodic hill geometries even when trained on just one or two DNS cases

Pope, two thousand: <ref:2605.26358#pg2>. They showed it can transfer from an attached duct flow to a separated hill flow without needing any retraining.

Lev: For us building the hardware side, the constant memory steady-state adjoint formulation is really what makes this practical. It requires the forward solver to converge to a steady state, which aligns perfectly with RANS because that’s what time-averaging means

Pope, two thousand: <ref:2605.26358#pg2>.

Kai: So for someone who just listens to the show, what does this mean? It means that for engineers and scientists relying on RANS simulations today, they have a new tool that can give them much more accurate results across a wider variety of real-world flows without needing to run incredibly expensive direct simulations.

Mira: It means we can use these physics-informed deep learning models to get better predictions where data is scarce, and those corrections stay robust even when the flow moves from one regime to another

Pope, two thousand: <ref:2605.26358#pg2>.

Lev: For real hardware deployment, the main thing is that this method works on two dimensional grids, which is a limitation you have to keep in mind when planning how you might use it for more complex three dimensional flows.

Kai: That’s right. We’ve seen how they combine a closure with physics-derived structure and a steady-state adjoint that provides solver-consistent gradients at constant memory

Pope, two thousand: <ref:2605.26358#pg2>. And that’s where this one ends for now.

Conclusion: Kai: So that's where we are on "Deep Learning-based Algebraic Reynolds Stress Closures for RANS Simulations of Turbulent Flows." Basically, they built this deep learning model to model the empirical constants of an algebraic Reynolds stress equation, and they trained it by optimizing over the whole system using discrete semi-implicit adjoint equations.

Mira: That's right. The big thing is that this method lets you train on a small dataset—like just one or two direct numerical simulations—and then it generalizes across different things like Reynolds number variations, unseen geometries, and even totally different flow regimes without needing to retrain the model

Pope, two thousand: .

Lev: And that constant memory gradient calculation is pretty efficient for real hardware. It means you don't need a massive amount of RAM just to train this model effectively on your system for solving those RANS problems

Pope, two thousand: .

Kai: I think the physics-derived structure is what really sells it. They aren't just throwing random numbers into a neural network; the tensor structure and the invariant basis come from first principles derived from the Reynolds stress transport equations

Pope, two thousand: .

Mira: That’s true. It gives it that inductive bias, which means it learns physical mappings instead of just guessing all the physics from scarce data

Pope, two thousand: .

Lev: The only caveat is what they mentioned: this approach relies on the weak-equilibrium assumption of their algebraic Reynolds stress model, so if you have flows with really strong curvature or rapid acceleration, it might fall outside its range of validity

Pope, two thousand: .

Kai: So, for someone listening who just wants to know what this means for engineering—it means we could get much more accurate turbulence predictions in RANS simulations across all those different conditions without having to run a million expensive direct simulations

Pope, two thousand: .

Mira: It’s about making the simulation tools accessible and reliable for situations where high-fidelity data is just not available

Pope, two thousand: .

Lev: From an error-correction standpoint, the hybrid discrete adjoint method is pretty solid. It reuses the fast numerical solvers for implicit steps when calculating those gradients, which keeps the overhead down on real hardware

Pope, two thousand: .

Kai: So that's it for this paper. It’s a solid step toward making these complex turbulence models more practical to use in everyday engineering simulations.

Mira: Definitely a significant advancement for how we model these turbulent flows. Next up, let's see what they're doing with those exotic electron-electron interactions and whether precision atomic spectroscopy can really probe physics beyond the standard model <ref:2602.09743#pg2>.

Lev: I'm curious to hear what kind of constraints they put on their numerical setup for that kind of experiment, though.

More episodes

← Home