Although molten salt reactor designs are attracting considerable attention as a fourth-generation technology with considerable advantages the technology represents challenges too.
The salt used to transport fuel around the reactor and provide heat transfer must possess specific characteristics. Reactor designers are looking for a combination of the right temperature, the right density, and viscosity, which are all related to how the liquid salt is going to be circulated and used to get the heat from the reactions out. There are still a lot of unknowns in the molten salt reactor area, like for example the different properties of salts of different compositions. Indeed, there are tens of thousands of different blends and combinations of salts that could be mixed at different proportions and used in reactor applications. This makes the task of choosing the ideal salt combination incredibly complex.
Now, to help pinpoint the right types of molten salts, researchers at the U.S. Department of Energy’s (DOEs) Argonne National Laboratory are deploying in-house developed artificial intelligence technology to explore the characteristics of molten salt combinations in a bid to identify the most promising for further research.
“The reason that we have different salt compositions and mix them is that they have got different properties. Another factor is that in the reactor environment, when we introduce uranium fuel this is a dynamic process affecting uranium concentrations because the process is prompting decay. Over time the uranium concentration will change and that will affect the salt properties as well. When you change the composition, the properties will change, like density, viscosity, heat capacity, which are all very important for the nuclear reactor operations,” explains Jicheng Guo, a Chemical Engineer at the Chemical and Fuel Cycle Technologies Division of Argonne National Laboratory.
Unravelling the liquid structure of multicomponent molten salts is challenging due to the difficulty of conducting high-temperature diffraction experiments and interpreting the results.
“Previously, if we perform experiments to explore these characteristics we have to prepare very many different compositions and look at the structure of each one under high energy X-ray, which is time consuming,” says Guo, adding: “Also, working with high temperature molten salt containing radioactive elements is very challenging and very costly because we have to have specialised staff to operate the instruments.”
However, another approach is to use computer modelling of the salt using different technologies.
Taking modelling to the next step
The traditional commercial technology for modelling these kinds of interactions is molecular dynamic simulation. There are two types of that simulation. One is classical molecular dynamics (MD) based on empirical forcefield, in which a very large cell containing a thousand to ten thousand atoms is modelled. “We can get a rough idea of what’s going on with the properties of this cell but it’s not so accurate because it uses some approximations,” says Guo.
Another approach is ab initio MD, which uses density function theory (DFT) to model a very small cell of maybe a hundred atoms. “The advantage of that is very accurate but it can only predict behaviours of a very small cell, so we don’t know what’s happening on the longer range,” notes Guo.
In a bid to address the problems associated with either of these approaches, researchers at Argonne developed a machine learning based Gaussian approximation potential (GAP) capable of learning many-body interactions directly from DFT*. It combines the advantages of these two computer simulations by simulating a very large cell with tens of thousands of atoms, but can still maintain the ab initio accuracy.
“This is a great advantage compared to both of these other techniques,” says Guo. Their research study set out to determine whether computer simulations driven by machine learning could guide and refine real-world experiments that are then conducted at the Advanced Photon Source (APS), a DOE Office of Science user facility.
Although initially modelled for specific molten lithium- potassium chloride (LiCl-KCl) salt combination, the key to the AI tool is the composition-transferable Gaussian approximation potential (GAP), as Guo says: “I think unique to this research is the ability to have a composition transferable potential, which means we can predict compositions that we didn’t introduce in the training set.”
While machine learning typically involves training a computer to analyse a situation based on existing data, in this case the researchers did not have validated examples that would normally be used by the machine model to learn. “One of the major problems in machine learning research is it doesn’t work beyond the training data set. This is a topic which many researchers in the general machine learning community are grappling with. How do you make sure these machine learning tools work in regions beyond the training data?,” says Ganesh Sivaraman, Assistant Computational Scientist at the Data Science and Learning Division at Argonne. “The problem here is how to incorporate all the training data for the mixtures where you have not trained the model,” he adds.
Building on previous modelling work, the researchers used active learning to create a transferable model to analyse molten salts. An accurate GAP was active-learned from only around 1100 training configurations drawn from 10 unique mixture compositions. This data was enriched with metadynamics.
“What machine learning can do is learn the underlying non-linearity, how do the atoms interact within the neighbourhood. That is learned from the training data set which would come from the most accurate data. We can solve the Schrödinger’s equation up to the physical limit of the chosen DFT approximation and essentially we do that for a very small group of atoms and then use that to train these machine learning models,” explains Sivaraman.
“Then the machine learning models can be used to run the same length scales, which is required for these modern solve simulations, but with the same accuracy as the first principles ab initio approach,” he says.
“We employ the partially trained machine learning model with a fixed number of compositions to force it to go to a region it has not seen using metadynamics. We break the potential multiple times to make sure that it captures all these different regions which is not explicitly included in the training data but then we incorporate those in the re-training process. With such a sampling it can interpolate from there an infinite number of combinations.
“Instead of changing the composition, we mimic all the scenarios which correspond to composition change and then we iteratively incorporate that into the machine learning model,” says Sivaraman.
This approach means that rather than being fitted for one or two specific molten salt mixture compositions, the transferable model can be applied to multiple mixtures.
The model is then able to make its predictions based on these principles, rather than a knowledge-based data set. Having developed the model the machine learning simulations were run using high performance computing resources at the Argonne Leadership Computing Facility (ALCF).
As Sivaraman says “We didn’t train the model with examples of that sweet spot (i.e., eutectic) composition, where you get the right melting point. Our model managed to predict that sweet spot, even without the corresponding training input.”
For the molten salt, while it is possible to mix them in different ratios and have almost infinite combinations of mixtures it is possible to use the model to explore promising new salt compositions.
“We show that you can indeed do that with machine learning models and then validate that data with physical experiments which makes science more rigorous. What this means is that you can discover salt compositions that are most relevant to the application, since the model can work in all those compositions even where you’ve not explicitly trained with actual data,” notes Sivaraman.
Researchers use the powerful X-rays at the APS to look closely at the structure of specific salt mixtures using high energy X-ray diffraction to validate the model. APS is a unique tool for these types of measurements and the facility’s 6-ID-D beamline was used to verify the AI analysis. “It’s typically very difficult for a simulation to match the readings from instruments, especially for high temperature models but when we compare the results that we get from this simulation to instrument readings from high energy diffraction it matches very well,” Guo noted, adding: “We used experimental results to validate our simulation. At the same time, the simulation results provided us more details about which salts to study further. They work with each other.” Guo adds: “This allows us to study multiple compositions at the same time.”
Increasing complexity
With the possible composition for reactor molten salts so enormous, it is almost impossibly expensive and time consuming to try to produce experimental data for every possible composition. The time and cost associated with real-world experiments make it desirable to narrow the field of candidates that undergo inspection so instead, the scientists turned to AI as the latest machine learning models can understand things they haven’t been shown and can extrapolate valuable data.
“We have already validated six different compositions where it was never trained. It works and we even have experimental validation for the eutectic, which is an important first result in this domain of machine learning,” says Sivaraman, who adds: “At the moment the research is focused on validating the model that has been devised. Then it will be deployed in a bid to identify desirable salt mixtures for use in reactors”.
Indeed, now that the researchers have shown this approach can work, the next step is to work with even more complex data that more closely reflects the actual environment of an operating molten salt reactor. “A molten salt reactor is quite a dynamic environment. The conditions change over time, and sometimes impurities can get into the salt,” Guo said. “We want to introduce a tiny amount of these impurities to see whether the model can predict how that affects the overall structure of molten salts and their properties.”
In addition when impurities go into the salt, it starts to corrode those high-temperature alloys, such the special alloy developed for the molten salt reactor called Hastelloy-N. These materials can also enter the salt. Furthermore, the model must also take into account the decay products as the reactor is operated and the uranium changes into different materials with different densities.
Once the scientists have fully validated the model, the goal is to be enter a series of desirable characteristics such as density or heat capacity and melting point and the model will come up with the correct composition.
“In the next step, we’re going to show we can pick random molten salts from the periodic table, mix them together and create models and it works really well. For some of them, there have not even been any experiments until now. That is the dream. Pick three or two different salts from the periodic table and then make a model for that and create conditions like the different melting points for that and figuring that out without doing a physical experiment. We have that capability now,” Sivaraman concludes.
The key to this project is not just finding a salt combination that has the desirable density and melting point characteristics, it’s determining how it is actually going to operate within a reactor system.
* GAP was first proposed and developed by Prof. Gábor Csányi at University of Cambridge