NSUF 21-4247: Coupling CFD and ML to transform the coated nuclear fuels fabrication process

Current Tristructural Isotropic (TRISO) fuel particle development is inherently slow and capital intensive as it relies almost entirely on empirical experience. In this work, computational fluid dynamics, machine learning, additive manufacturing, and ORNL’s unique 50 years of experience in TRISO fuel particle development will be combined to transform the process of optimization and design of fluidized-bed chemical vapor deposition (FBCVD) systems for the production of current and future TRISO-like fuel particles. Machine learning algorithms will actively explore and recommend promising candidates for new FBCVD configurations as the model is constantly refined using a Bayesian approach for validation and calibration using the ORNL-developed Tasmanian library. Computational fluid dynamics (CFD), using the DOE-developed Multiphase Flow with Interphase eXchanges (MFiX) software, will be used to feed the machine learning (ML) algorithm by designing new gas distributors and reaction chambers, and measure its effect on the movement of each of the ~130k surrogate kernels and on the overall gas distribution in a 50 mm diameter FBCVD coater. The trained ML surrogate model will be used to select three optimal geometries and associated process conditions, and additional CFD simulations will be performed using these parameters in order to validate the ML model predictions. This ML-directed active search for new designs will drastically reduce the number of experiments required for the development and optimization of FBCVD systems for current and future nuclear and other energy related applications. The project outcome would be a novel, integrated process capable of testing a larger amount of unexplored designs and combination of variables that would be impossible to achieve purely using experimental routes and traditional machining technologies. The impact would be a flexible and faster implementation of new FBCVD coaters for the production of current and future fuels via the production of a transformational modeling and simulation platform. Since ML algorithms are trained on datasets to learn, create and predict new properties, the creation of the first dataset on FBCVD for nuclear fuel production will present a significant opportunity to train and validate new machine learning models that will be required during future fuel development.

Additional Info

Field	Value
Abstract	Current Tristructural Isotropic (TRISO) fuel particle development is inherently slow and capital intensive as it relies almost entirely on empirical experience. In this work, computational fluid dynamics, machine learning, additive manufacturing, and ORNL’s unique 50 years of experience in TRISO fuel particle development will be combined to transform the process of optimization and design of fluidized-bed chemical vapor deposition (FBCVD) systems for the production of current and future TRISO-like fuel particles. Machine learning algorithms will actively explore and recommend promising candidates for new FBCVD configurations as the model is constantly refined using a Bayesian approach for validation and calibration using the ORNL-developed Tasmanian library. Computational fluid dynamics (CFD), using the DOE-developed Multiphase Flow with Interphase eXchanges (MFiX) software, will be used to feed the machine learning (ML) algorithm by designing new gas distributors and reaction chambers, and measure its effect on the movement of each of the ~130k surrogate kernels and on the overall gas distribution in a 50 mm diameter FBCVD coater. The trained ML surrogate model will be used to select three optimal geometries and associated process conditions, and additional CFD simulations will be performed using these parameters in order to validate the ML model predictions. This ML-directed active search for new designs will drastically reduce the number of experiments required for the development and optimization of FBCVD systems for current and future nuclear and other energy related applications. The project outcome would be a novel, integrated process capable of testing a larger amount of unexplored designs and combination of variables that would be impossible to achieve purely using experimental routes and traditional machining technologies. The impact would be a flexible and faster implementation of new FBCVD coaters for the production of current and future fuels via the production of a transformational modeling and simulation platform. Since ML algorithms are trained on datasets to learn, create and predict new properties, the creation of the first dataset on FBCVD for nuclear fuel production will present a significant opportunity to train and validate new machine learning models that will be required during future fuel development.
Award Announced Date	2021-06-07T16:21:17.84
Awarded Institution	Idaho National Laboratory
Facility	Advanced Test Reactor
Facility Tech Lead	Alina Zackrone
Irradiation Facility	None
PI	Zachary Mills
PI Email	[email protected]
Project Type	RTE
RTE Number	4247