An Invariant Neural-Network Eddy-Viscosity Closure with Iterative Solver Coupling for Molten Salt Fast Reactor Thermal-Hydraulics
Authors: Zaid Abulawi, William Oliveira, Akzhol Almukhametov, Jun Fang, Ling Zou, Rui Hu, Stefano Lorenzi, Yang Liu
Organizations: Department of Nuclear Engineering, Texas A&M University · Department of Nuclear Engineering, Texas A&M University, College Station, TX 77843, USA · Department of Energy, Nuclear Engineering Division, Politecnico di Milano · Department of Energy, Nuclear Engineering Division, Politecnico di Milano, Milan, Italy · Nuclear Science and Engineering Division, Argonne National Laboratory · Nuclear Science and Engineering Division, Argonne National Laboratory, Lemont, IL 60439, USA
System thermal-hydraulics codes such as the System Analysis Module (SAM) close turbulence on coarse meshes with algebraic zero-equation mixing-length models. These models cannot represent the non-equilibrium turbulence, flow separation, and large-scale recirculation of open-cavity advanced reactors. We develop an invariant neural-network eddy-viscosity closure for SAM and apply it to the Molten Salt Fast Reactor (MSFR). The closure predicts the turbulent kinetic energy and timescale of each element from the flow and geometry of its neighbors, without solving turbulence transport equations. The resulting non-negative eddy viscosity enters the implicit diffusive operator, which keeps the momentum equations well conditioned. We prove that the closure is exactly invariant to translation, Galilean boost, rotation, reflection, and neighbor ordering. A network trained on high-order Nek5000 fields sees a different feature distribution on the coarse SAM mesh. We remove this covariate shift by iteratively retraining the network on SAM flow states against fixed Nek5000 targets. For a held-out steady state, the coupled closure reproduces the single recirculation cell and eddy-viscosity plume of the Nek5000 reference. In a 120 s pump-coastdown transient, it halves the velocity error of the mixing-length model, from 37.6% to 17.7%, and matches the Nek5000 eddy viscosity with a time-averaged relative error of 12%. The remaining velocity error comes from the coarse mesh. The coupled transient takes about 0.2 core-hours, compared with about 330 core-hours for Nek5000. The results show that the closure can recover two-equation RANS turbulence behavior in a coarse-mesh system code at a small fraction of the CFD cost.
Figures & tables
Figure 1: Overview of the closure and its coupling with SAM. (a) The eddy viscosity μt at each target element ( ∗ ) is inferred from the flow state on its n neighboring elements ( ∙ ). (b) A shared-weight embedding network maps the scalar features of each row of Q∈Rn×9 to the embedding matrix G . (c) The invariant descriptor D (Eq. ( 15 )) and Re are mapped to k and τ , giving μt=ρkτ≥0 . (d) μt is evaluated once per time step from the accepted state, capped at 30 Pa s, and held fixed during the Newton iterations; a failed step is retried with a smaller Δt .
Figure 2: Iterative training. (a) Training strategies on the coarse mesh: no neural closure, offline training on Nek5000 features, and retraining on SAM states. (b) Iterative training: each coupled SAM run adds its feature states to the aggregated dataset A , and the closure is retrained on A against the fixed Nek5000 targets without differentiating through the solver.
Figure 3: Computational meshes and boundary conditions for the axisymmetric MSFR cavity. a , Nek5000 spectral-element mesh ( 4,456 elements, p=7 ). b , Coarse SAM mesh ( 1,296 elements). c , d , Near-wall views of the dashed region in Nek5000 (eight boundary-layer elements) and SAM (two).
Parameter
Case 1
Case 2
Case 3
Case 4
Case 5
Case 6 (Test)
Reynolds number, Re
2.26×106
1.50×106
1.13×106
9.03×105
5.64×105
4.52×105
Inlet speed, Uin (m/s)
1.25
1.25
1.25
1.25
1.25
1.25
Length scale, L0 (m)
2.0
2.0
2.0
2.0
2.0
2.0
Density, ρ (kg/m 3 )
4125
4125
4125
4125
4125
4125
Viscosity, μ (Pa s)
0.0046
0.0069
0.0091
0.0114
0.0183
0.0228
Table 1: Operating parameters and thermophysical properties for the six steady-state MSFR cases.
Quantity
Value
Fluid density, ρ
4306.71 kg/m 3
Molecular dynamic viscosity, μ
0.00894 Pa s
Molecular kinematic viscosity, ν
2.08×10−6 m 2 /s
Reference CFD turbulence closure
Two-equation k – τ RANS
Initial peak inlet velocity, Upk(0)
1.25 m/s
Asymptotic peak inlet velocity, U∞
0.30 m/s
Table 2: Thermophysical properties and boundary-condition parameters for the 120 s transient flow-decay benchmark.
Parameter / Component
MSFR Implementation Specification
Characteristic scales (L0,U0,T0)
0.235 m ( 2.0 m steady-state), 1.25 m/s, 0.188 s
Strain-rate normalization in ψc , Eq. ( 12 )
21lnmax(∥Sa∥FT0,10−12)+2
Relative-speed normalization in ψc , Eq. ( 12 )
61ln(1+100max(∥wa∥2,10−8))
Wall-distance normalization in ψc , Eq. ( 12 )
min(dw,a/L0,1)
Target k transform in ψz , Eq. ( 17 )
zk=21ln(1+1250k/U02)
Target τ transform in ψz , Eq. ( 17 )
zτ=41[lnmax(τ/T0,10−12)+7]
Table 3: Case-specific normalization maps and neural-network architecture hyperparameters for the MSFR closure implementation.
Figure 4: A priori parity plots of the network predictions against the Nek5000 k – τ reference. Columns: k , τ , and μt . Rows: early extrapolation ( 2 – 10 s; a – c ), interior interpolation ( 20,40,60 s; d – f ), and late extrapolation ( 70 – 120 s; g – i ). Dashed lines indicate 1:1 agreement.
Figure 5: Temporal evolution of a priori prediction errors across t∈[2,120] s. a – c , Relative L2 errors for k ( a ), τ ( b ), and μt ( c ). d – f , Physical-unit RMSE for k ( d ), τ ( e ), and μt ( f ). The shaded amber band denotes the 11 – 69 s training window, within which t=20,40,60 s are held out for interior interpolation testing.
Figure 6: Steady-state a posteriori comparison for held-out Case 6 ( Re=4.52×105 ): Nek5000 + k – τ (left) and SAM + NN (right). Rows: axial velocity uz (labeled ux ), negative radial velocity −ur (labeled uy ), and eddy viscosity μt . Panel axes x and y correspond to r and z . The symmetry axis is at the left edge and the cold-leg inlet at the bottom, as in all contour figures.
Figure 7: Velocity magnitude ∣u∣ at t=10,30,60, and 120 s (rows). Columns: projected Nek5000 reference, SAM + NN, SAM + mixing length, and SAM with prescribed Nek5000 μt ; panels a – p are ordered row by row. In-panel labels give the relative L2 velocity error.
Figure 8: A posteriori transient eddy-viscosity fields μt at t=10,30,60, and 120 s (rows) comparing the projected Nek5000 + k – τ reference ( a , c , e , g ) and SAM + NN ( b , d , f , h ) on the 1,296-element coarse mesh. In-panel labels give the volume-weighted relative L2 eddy-viscosity error.
Figure 9: One-dimensional transient profiles of velocity magnitude ∣u∣ (left column) and eddy viscosity μt (right column) along the cross-core sampling line r=0.50 m (indicated in the top-left inset) at t=10,30,60, and 120 s.
Checkpoint / Temporal Statistic
Velocity Error Eu (%)
Viscosity Error Eμt (%)
t=10 s (early extrapolation)
17.88
11.28
t=30 s (training window)
20.09
4.85
t=60 s (interior held-out checkpoint)
18.90
4.51
t=120 s (end of late extrapolation)
14.08
29.71
Time-averaged error Eq ( 2 – 120 s)
17.73
12.02
Table 4: Volume-weighted relative L2 errors of SAM + NN relative to the projected Nek5000 reference across the 120 s transient, evaluated via Eq. ( 28 ).
Figure 10: Temporal evolution of a posteriori volume-weighted relative L2 errors for SAM + NN across t∈[2,120] s. a , Velocity error Eu(t) ; the dotted line is SAM with prescribed Nek5000 μt . b , Eddy-viscosity relative error Eμt(t) . Dashed horizontal lines denote the time means ( 17.73% and 12.02% ), and orange markers highlight the four snapshot times displayed in Figs. 7 – 9 .
Figure 11: Convergence of the iterative training at t=30 s. a , Predicted velocity magnitude ∣u∣ (top row) and eddy viscosity μt (bottom row) comparing the Nek5000 + k – τ reference against coupled SAM + NN Iterations 0–6 (Iteration 0: network trained on Nek5000 features; Iterations 1–6: retrained on all previous SAM runs). b , Corresponding pointwise absolute discrepancy fields ∥Δu∥2=∥u−u∗∥2 (top row) and ∣Δμt∣=∣μt−μt∗∣ (bottom row), where the inset labels report the domain-integrated volume-weighted relative L2 error for each iteration.
Figure 12: Convergence of the iterative training for the transient. Time-averaged ( 2 – 120 s) volume-weighted relative L2 errors of a , the velocity vector and b , the eddy viscosity versus iteration, for retraining on the latest run (orange) and on all previous runs (blue). The dotted line in a is SAM with prescribed Nek5000 μt .
Coarse mesh
Fine mesh
First order
Second order
Third order
First order
Elements
1,296
1,296
1,296
4,456
Degrees of freedom
4,197
16,167
35,913
13,926
Velocity error Eu at the coarse element centers (%)
t=10 s
20.02
20.11
19.56
7.04
t=30 s
19.61
19.03
18.53
5.19
Table 5: Velocity error of SAM with prescribed Nek5000 μt , using first-, second-, and third-order velocity elements on the coarse mesh and first-order elements on the fine mesh. Eu is the volume-weighted relative L2 error of Eq. ( 28 ).
Simulation
Code
CPU cores
Time steps
Wall-clock time
CPU core-hours
Steady state
Nek5000 + k – τ
64
3.6×106
22.7 h
1,450
Steady state
SAM + mixing length
1
64
2.3 min
0.038
Steady state
SAM + NN
1
63
4.5 min
0.075
Transient (120 s)
Nek5000 + k – τ
32
1.2×106
10.3 h
330
Transient (120 s)
SAM + mixing length
1
169
6.3 min
0.11
Transient (120 s)
SAM + NN
1
169
10.5 min
0.17
Table 6: Computational cost of the Nek5000 reference simulations and of SAM coupled with the mixing-length model and with the final neural closure.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Closure
Coupling
Steps
Newton its./step
Failed solves
Residual evals.
Transient (0–120 s)
Mixing length
Implicit
169
2.67
0
19,692
Neural closure (final)
Lagged
169
2.60
0
19,176
Prescribed Nek5000 μt
None
169
2.24
0
16,510
Steady state (pseudo-transient)
Mixing length
Implicit
64
2.48
0
6,907
Appendix
Table 7: Convergence of SAM with three eddy-viscosity closures, with identical mesh, boundary conditions, and solver settings.
A neural turbulence closure defines a new boundary-value problem, R(U)=N(U)+F(U)=0, with a coupled Jacobian J(U)=N′(U)+F′(U), where N is the original mean-flow operator and F the learned closure. We establish two consequences of global tangent dissipation. For a monotone original operator, a positive uniform margin supplied by the original operator and closure together guarantees existence, uniqueness and a global inverse-sensitivity bound relating a posteriori solution error to the a priori residual. For a general original operator, a dissipative closure cannot worsen tangent dissipation, but this alone does not guarantee uniqueness. Tangent dissipation depends on both diffusion and reaction. We study two complementary ways to promote it: (1) an exact-integral construction enforcing non-negative tangent diffusion while leaving reaction unconstrained, and (2) a penalty on tangent-reaction violations at sampled states. Tangent diffusion enters the Jacobian, and non-negative secant eddy viscosity alone does not control its coercivity. We conduct tests with channel flow at Reτ=180--5200, which provides a strongly monotone baseline. Both constrained closures reach accurate solutions in all 50 training-seed/Reynolds-number cases. At Reτ=1000, we conduct tests with 10,000 starts for one fixed network per closure and we find one root for each constrained closure and multiple roots for the other closures. Although this does not prove uniqueness, it provides strong empirical evidence for uniqueness of the tested constrained closures. At Reτ=5200, the construction and penalty reduce the reported inverse sensitivity relative to the original operator by approximately 372× and 11×, respectively.
Zhen Zhang, George Em Karniadakis
Division of Applied Mathematics, Brown University, Providence, RI 02912, USA
Turbulence is ubiquitous in engineering and science, yet direct simulation is prohibitively expensive. The Reynolds-averaged Navier-Stokes (RANS) equations provide savings exceeding ten orders of magnitude but introduce unclosed terms. Offline-trained machine-learning (ML) closures suffer distribution shift 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 for RANS, the Deep Algebraic Reynolds Stress Model (DARSM), which can be trained on one or two flow cases and accurately generalise across an order of magnitude in Reynolds number, to unseen geometries, and across 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 physics-based structure on the ML closure. End-to-end optimisation through the governing PDEs eliminates distribution shift, but both unrolled and implicit automatic differentiation fail on the stiff coupled solver. We derive adjoint equations that exploit the solver's implicit-explicit structure for efficient optimisation. On canonical square-duct and periodic-hill benchmarks, DARSM reduces average test velocity error over baseline RANS by 2-4x across Reynolds number, geometries, and flow regimes, with peak case-level reductions of 12x. The model trained on attached, anisotropy-dominated flows (square duct) accurately generalises without retraining to separated flows (periodic hills), a regime change in the underlying physics. DARSM also outperforms established ML methods: offline training, tensor-basis neural networks, field-inversion machine learning, DeepONets, physics-informed neural networks, and an adjoint-trained source-term closure.
Daniel Dehtyriov, Jonathan F. MacArt, Justin Sirignano
Mathematical Institute, University of Oxford, Oxford OX2 6GG, United Kingdom · Aerospace and Mechanical Engineering, University of Notre Dame, Notre Dame, IN 46556, USA
The differentiable physics paradigm may be leveraged as an a-posteriori approach for discovering turbulence closure models by embedding a neural network parameterization directly inside the solver and optimizing it given potentially sparse target data. This addresses a key limitation of a-priori learning where direct numerical simulation (DNS) data is used to approximate the subgrid stress with the assumption of a low-pass filter. Closures trained in this a-priori manner frequently lead to unstable deployments due to the mismatch between the assumed filter and the effect of numerical discretizations and coarse-graining. In comparison, while typically stable during deployment, a-posteriori learning incurs high computational costs due to the need to backpropagate through a large eddy simulation (LES) solver. Furthermore, a-posteriori methods are challenging to apply broadly since they require significant modification of existing solvers. Finally, both approaches are limited when generalization is desired across different numerical schemes with their implicit filtering characteristics. In this work, we present a deep-learning approach for turbulence closure modeling built on the continuous data assimilation framework. Our approach enables the a-priori training of closures using sparsely observed DNS data without modifying or differentiating through the LES solver, while preserving stability during deployment for the recovery of invariant statistics. We focus on the model's ability to adapt to different discretizations by explicitly conditioning it on the numerical scheme. We use two- and three-dimensional canonical cases to test our framework and show that the learned correction systematically tracks the discretization error of the coarse solver.
School of Mechanical Engineering Purdue University West Lafayette, IN 47907 · College of Information Sciences and Technology Pennsylvania State University University Park, PA 16802