Research

SMR Components & Design Concepts

We work on the dynamic loads on the emergency depressurization and recirculation valves of the i-SMR, the flow characteristics inside and outside helical coil tubes, and two-phase flow.

i-SMREmergency valve dynamic loadsHelical coil tubeTwo-phase flowThermal-hydraulic design

Concept

Computational Fluid Dynamics & Validation

We compute velocity and temperature fields where measurement is impractical, and establish how far those numbers hold by putting them against experiment. The longest thread runs through wire-wrapped fuel bundles for sodium-cooled fast reactors.

RANS · URANSLES · DESWire-wrapped bundlesSubchannel flow distributionValidation against measurementGrid convergence (GCI) & uncertainty quantificationAnsys · OpenFOAM · STAR-CCM+Core-scale methodology

Concept

Flow Visualization Experiments

We match the refractive index of the working fluid to a transparent model of a reactor internal, making the walls optically vanish, and measure the velocity field inside. Those measurements become the reference against which simulations are checked.

Matched index of refraction (MIR)Particle image velocimetry (PIV)Planar laser-induced fluorescence (PLIF)Magnetic resonance velocimetry (MRV)SLA 3D-printed test sectionsCFD validation data

Concept

Passive Safety Systems & Natural Circulation

Heat has to leave even with no power and no pump. We work on three routes: natural circulation driven by density difference, heat pipes that carry heat by phase change, and geometry that sets a direction with no moving part.

Natural circulationExternal reactor vessel cooling (ERVC)Passive fluidic diodePassive safety and depressurization valvesHeat pipes and vapour chambersSystem codes (MARS-KS)

Concept

Additively Manufactured Heat Exchanger Design

We design heat exchangers that are smaller yet move more heat, using flow-passage geometries only additive manufacturing can produce.

Additive manufacturingTPMS structuresCompact heat exchangersPressure loss vs. heat transferManufacturability

Concept

Machine Learning for Thermal Hydraulics

We cut the cost of simulations that CFD takes far too long to run, and recover the three-dimensional information two cameras cannot see. Only validated simulation and experimental data goes into training.

Reduced-order modelsPOD · LSTM3D convolutional networksPhysics-informed neural networksTraining dataset generationWorld models

Concept

Probabilistic Safety Assessment for Advanced Reactors

For reactor types that have no established framework yet, such as the molten salt reactor, we work out which accident paths exist and how often they can occur. Rather than stopping at counting branches, we run the analysis codes themselves so that timing comes with the answer.

Probabilistic safety assessmentDynamic PSAMolten salt reactorFloating reactorEvent and fault treesState-based modelingRAVEN · EMRALDExternal events (seismic, flooding)

Concept

Spectral Element Method for Large-Scale Reactor Analysis

Resolving a domain as large as a whole core is not a matter of simply making the mesh finer. We raise the order of the solution inside each element instead.

Spectral element methodNekRSHigh-order accuracyLarge-scale parallel computingHelical coil rod bundlesCore-scale analysisLarge eddy simulation

Where we are headed — not something we have done

A model that rolls a whole reactor forward

Today we teach a model one flow, one geometry at a time. What we actually want is a model that has learned how a reactor behaves as a whole, one you can hand a changed condition and let it roll forward on its own. A world model, in other words.

What we have: We generate training data from high-fidelity simulation, compress the flow into a few dozen representative shapes, and have a network carry it forward one step at a time. For cross flow between steam generator tubes, a 33-hour computation came down to 35.5 seconds. We have also trained with physics as a constraint, and recovered three-dimensional shapes from two images.
What is missing: A model of that kind is right only inside the geometry and the conditions it was trained on. Change the tube pitch or raise the power and it has to learn again from scratch. People do not work that way: once the feel is there, a layout you have never seen is still readable. Giving a machine that feel is what remains.
Why it is worth wanting: Design is, in the end, asking “what if we change this” a few thousand times. When one such question costs a day, the number of questions is what limits you. Make the rolling-forward cheap and the limit becomes the quality of the questions instead. Playing an accident sequence forward before it happens is the same problem.

There is a long way to go. The reduced-order models and physics-constrained learning we are building now are the footing for it.

See machine learning for thermal-hydraulics