koroush shirvan john clark hardwick (1986) career

17
Application of Studsvik Tools to LWR Core Optimization and Advanced Fuels Engineering Koroush Shirvan John Clark Hardwick (1986) Career Development Professor Sept 23 rd , 2021

Upload: others

Post on 13-Mar-2022

6 views

Category:

Documents


0 download

TRANSCRIPT

Application of Studsvik Tools to LWR Core Optimization and Advanced Fuels Engineering

Koroush ShirvanJohn Clark Hardwick (1986) Career

Development ProfessorSept 23rd, 2021

About Me• 30+ researchers in NIFT (established

2017) Mostly Graduate Students

• Focus: Integration of disciplines to develop and understand innovative nuclear technologies and their performance.

• Director of Reactor Technology Coursefor Executives (Former MIT CASLEducation Chair)

• Raised >$30 million last 5 years Including Industry: EDF, General

Atomics, FORTUM, Exelon, Lockheed Martin, Exxon

Economics& Licensing

6

Thermal Hyd.& Safety

9

Reactor Physics

8

Structural Mechanics

8

System-Level Integration

Center for Nuclear Innovation in Fission Technology [NIFT]

PI: Koroush Shirvan

2

MIT Use of STUDSVIK Tools• 2008-Present: C4/S3 – Small Modular Reactor (SMR) Core Design

• 2010-12: C4/S3 /S3K – BWR Assembly and Core Optimization, Stability Analysis

• 2012-14: C4/S3 – S3 vs. full core 2D C4e to Watts Bar plant detector data

• 2012-16: C4/S3/S3K– Advanced Fuel Designs including Thoria-Urania-Plutonia, UO2-BeO, Accident Tolerant Fuels (SiC/Coated/FeCrAl, UN/ U3Si2 Fuels)

• 2013-Present: C4 – Advanced Reactors (Superheated Reactor, Organic Cooled Reactor, Thermal Spectrum Micro-Reactors, CANDU Thoria fuel radial power)

• 2017-Present: S3K – Post-CHF Simulations to support time/temperature criterion

• 2019-Present: C4/S3 – AI Driven PWR/BWR Core/Fuel Optimization [Exelon]

• 2020-Present: C4/S3 – High Burnup/LEU+ (~5-8% Enrichment) Simulations

3

Why We Choose Studsvik Package (From My Lecture Slides)

MCNP5 ENDF7 0SERPENT ENDF7 -46SERPENT ENDF6 -166SERPENT JEFF2 -128

CASMO4e ENDF6 -640CASMO5 ENDF7* -42

NEWT ENDF7 -423NEWT ENDF6 -779

KENO-C ENDF7 +38KENO ENDF7 -516

DRAGON -ENDF7 -52

* The units are in PCM difference from MCNP5 ENDF7* CASMO models physics not existent in MCNP5 (C5 Simulation turned off these models)* All Monte Carlo values are less than 0.01% delta K/K * KENO-C = KENO-VI with Continuous energy library* Mesh convergence not achievable in NEWT

CASMO4 NEWT KENO DRAGON MCNP5/X SERPENT5 sec 8 minutes 17 minutes 40 sec 2.5 days 20 hours

*Single 2.6 GHz node on Linux operating machine

4

Advanced Fuel Design Requires Optimization

5

Impact of Fuel on Probability of Failure SiC-cladded fuel performance during normal operation

He Y., et al.,

Topfuel, 2019

Prob

abili

ty o

f Lea

king

C4/S3 ~109 x

faster than CASL Tools

FRPACON-MIT

~105 x faster than BISON

6

Value of Best Estimate • Our analysis showed noticeable gain

in margin when performing best estimate Rod Ejection Analysis (REA) for selected advanced fuels using S3K Studsvik tool. similar gains to best estimate

LOCA

• Traditional best estimate REA analysis involves sensitivity analysis & uncertainty quantification Requires fast running, multi-

physics, pin-resolved ROBUST reactor transient simulator package

REA Simulation for 4L-W PWR with Studsvik’s S3K

7

AI-Driven Nuclear Fuel Reload Optimization with C4/S3• Why Use ML/AI Instead of Other Algorithms?

50! (~3 x 1064) Possibility for Fuel Patterns Hard to explain problems: if it was easy, we would have

gotten the optimum solution Challenge: Enterprise Risk of Defective Nuclear Fuel is

Very High: < 5 Thousands of Percent Solution: C4/S3• Expert Knowledge is Limited

• Contracting different fuel vendor (Different fuel products)

• Change of Cycle Length (Market pressure may shorten or lengthen cycle length)

• Greater than 5% Enrichment and Higher Burnups (Lower front/backend fuel cycle cost)

Motivates Development of Physics-Based

Continual Learning & Efficient

Optimization Routine

8

Optimization with Reinforcement Learning (RL)

Radaideh & Shirvan, Knowledge-Based Systems, 2021https://doi.org/10.1016/j.knosys.2021.106836

Expert Rules for Successful Fuel Design

Incorporation of physics-based game tactics was key to the success of AI

Currently undergoing testing on Exelon Machines for few PWR and BWR Plants usingC4/S3 as Physics Packages

9

Why Not AI/ML for the Physics?!

Training Prediction (Unseen Data)

https://arxiv.org/abs/2104.09499

Current Implementation

in S3

10

High Burnup Fuel Impact• As burnup increases beyond

current limit of 62 MWD/kgUpeak average rod, the fuel performance feedback becomes stronger Plenum Pressure AO, CIPS, CILC Requires ROBUST multi-

physics fast running tool package to perform optimization

11

Shirvan K., TopFuel, 2021

Final Remarks• Speed and Accuracy provided by Studsvik tools are critical for their use for

Education and Research at Universities Ease in licensing tremendously helps with use in university environment. Compatibility with Windows/Linux (several architectures) Major downside relative to other available packages is the yearly fee

• MIT Reactor Design Research Groups have utilized CASMO4/SIMULATE3 package extensively mainly for Light Water Reactor Applications

• As industry moves to higher burnup cores, ATFs and new operational modes, more automation and tighter multi-physics coupling will be needed.

12

Backup

13

Shirvan Research Experience and Capability

Category Industry/ Auditing

Advanced M&S

Reactor Physics

STUDSVIK/PARCS SERPENT

Thermal-Hydraulics

VIPRE STARCCM

COBRA-TF

Safety/System RELAP5/ TRACE --

Fuel Performance

ABAQUS/ FRAPCON

BISONIn-House

Severe Accident

MAAP/ MELCOR --

Economics EEDB In-House

Experimental FacilitiesCategory Facilities (incl. Shared)

Corrosion 1500oC Steam*

High Pressure

Autoclave*

Thermal-Hydraulics

Quench Facility*

SEM/EDS/FIB/XPS

Strength/ Fatigue

Instron/4-point bend*

Burst/Plug*

Irradiation PSFC Ion Beam*

MIT Reactor*

Post Irradiation Examination

Furnace/ Instron* SEM/EDS*

*Capable of handling rad materials

Modeling and Simulation Network/Funding Profolio

Ability to Integrate Disciplines to Accelerate R&D and Explore Cost Reduction Innovations

14

Annual Cost of Existing U.S Nuclear Plants

• ~$70 million in regulation cost and half in NRC fees [American Action Forum Study]

• ~$5-10 million loss per year can result in shutdown and loss of >1GWe of carbon free electricity Adding 25 million tons of CO2

annuallyAccident Tolerant

Fuel

Digital Twins

AI

Modeling and Simulation

Opportunities

15

Replace Physics with AI/ML (1)

Training Prediction (Unseen Data)

https://arxiv.org/abs/2104.09499

16

Replace Physics with AI/ML (2)

Surrogate Training time [s](training size 9128)

Runtime [µs](on a single rod)

Runtime Acceleration(FRAPCON runtime / surrogate

runtime)LUT - 186.910 3.103E+04PLS 0.0288 0.102 5.686E+07SVR 224.867 128.082 4.528E+04

*SVR 0.0318 5.636 1.029E+06GP 930.980 1778.278 3.262E+03

*GP 4.233 167.316 3.466E+04NN-1layer 8.361 14.333 4.047E+05NN-2layer 10.496 16.389 3.539E+05NN-3layer 12.351 18.821 3.082E+05

RF 67.915 95.513 6.072E+04XGB 36.077 12.138 4.778E+05

**Combined surrogate ~ 63.160 472.277 1.228E+04

*GP and *SVR are trained on a smaller training set with 300 samples from the space-filling design**The combined surrogate composes of the best surrogates for prediction of all QoIs at the same time.

https://arxiv.org/abs/2104.09499

17