Ultra-Fast Modeling of Heat Treatment furnace. Advanced online control.

Boutros Ghannam

Maroun NEMER

Center for Energy Efficiency of Systems (CES) - Mines ParisTech

Thursday, May 04, 2017

60 Boulevard Saint Michel, 75014 Paris | France

Thermal numerical simulation, some examples

5/4/2017 2

iSGTW: A flame held by a flameholder. The FOCUS project simulations have demonstrated that taking radiation into consideration when making calculations modifies the dynamics of the flame. Time: 300000 hours and 400 CPUs Image courtesy of FOCUS

Steam explosion with nuclear fuel The radiativeheat transfer between the hot molten UO2 and the surrounding water is a key mechanism for controlling the boiling process Image courtesy of Walter W. Yuen

Steel Slabs Reheating Funace Coupled Radiation and heat diffusion Control strategies allow less energy consumption and better heating quality / Nevertheless their accuracy is very sensitive to the speed and the precision of the numerical simulations

Steel Slabs Reheating Furnace

5/4/2017 3

Computing 3D radiation heat Transfer

Computing heat diffusion in the slabs

Radiation heat flux B.C.

Coupling

3DFTDT

3D Temperature Profile in the slab

Direct exchange factors

Total exchange factors

Context and objectives of the thesis work

Calculation: TFLOPS

Graphics Processing Unit (GPU) Computing (1/2)

5/4/2017 4

CPU smoke simulation

GPU smoke simulation

Comparison of CPU and GPU performances

CPU and GPU peak performance Calculation : Tflops vs. 100 Gflops Memory bandwidth: ~10x

Graphics Processing Unit (GPU) Computing (2/2)

5/4/2017 5

GPU Architecture

CPU Architecture

GTX 680

Comparison of CPU and GPU architectures

Discretization

5/4/2017 6

The Multiple Absorption Coefficient Zonal Method (MACZM)

Volume elements (voxels) :

Uniform radiative properties (mean value)

Uniform temperature

GEF

Continuous space Voxelized space

Computation acceleration by Artificial Neural Networks

5/4/2017 7

The Multiple Absorption Coefficient Zonal Method (MACZM)

Artificial Neural Networks (ANNs) replace the integral computation for GEFs

Generic Exchange Factors (GEFs) Artificial Neural Networks (ANNs)

Application to a test furnace

5/4/2017 8

Numerical and experimental verification Comparison MACZM – MODRAY

Direct exchange factors

[Computation time]

MACZM

[2 sec]

MODRAY (a)

[25 min]

MODRAY (c)

[2 min]

Combustion volume(1)-

combustion volume(2) 1.13560E-04 9.98015E-05 1.27992E-02

Combustion volume(1)-

steel slab 1.03177E-02 1.00157E-02 8.68062E-02

Combustion volume(2)-

steel slab 1.04588E-02 1.01281E-02 8.56366E-02

Base - base 3.28529E-02 3.09024E-02 3.09022E-02 External view Internal view

Numerical validation

Experimental validation

Supports

Test Furnace mesh

Implementation of MACZM on the GPU

5/4/2017 9

Parallelization and device code optimization

GPU parallelization SIMD mode Each thread computes one GEF

(discrete line and ANN)

(Device) Grid

Constant Memory

Global Memory

Block (0, 0)

Shared Memory

Local Memory

Thread (0, 0)

Registers

Local Memory

Thread (1, 0)

Registers

Block (1, 0)

Shared Memory

Local Memory

Thread (0, 0)

Registers

Local Memory

Thread (1, 0)

Registers

Host Scene data ANNs

Discrete line algorithm Parametric 6-line (No-

divergence) Scene data in char Performance enhancement

using registers

Artificial Neural Networks Constant memory matrix

multiplication Synchronized call for

memory access efficiency

temporary data

3D Finite difference method

5/4/2017 10

Discretization and Differentiating schemes

Steel slab

𝝏𝑻𝝏𝒕= 𝛼

𝝏𝟐𝑻

𝝏𝒙𝟐𝐷 +

𝝏𝟐𝑻

𝝏𝒚𝟐+

𝝏𝟐𝑻

𝝏𝒛𝟐 , 𝛼 = 𝑘/𝜌𝑐𝑝

Heat diffusion PDE (3D):

B.C.: Radiation heat flux

Finite difference discretization schemes (1st order time - 2nd order space) Simple explicit : Forward time – centered space

Stability condition Accuracy : O(Δt)

Simple implicit : Backward time – centered space Unconditionally stable Accuracy : O(Δt) N3 x N3 matrix

Crank-Nicolson scheme: ½ explicit & ½ implicit Unconditionally stable Accuracy : O(Δt2) N3 x N3 matrix

Application to the simulation of a slab

5/4/2017 11

3000 simulations/second for 32 x 64 x 16 meshes

Thermal model

5/4/2017 12

Wall temperature:

Constant in vertical sections

Input for the heat balance

Rail Vertical section

Gas temperature supposed constant in vertical sections

Rails Cooled by water flow 2D analytical model (steady state)

Radiation flux is projected on the slabs for computing heat diffusion in the slabs

Industrial Solution

5/4/2017 13

Industrial Solution

5/4/2017 14

5/4/2017 © Boutros GHANNAM | Center for Energy and Processes - Mines ParisTech 15

Thank you