Generative
Deep Generative Learning
The confluence of ultrafast computers with large memory, rapid progress in AI/ML algorithms, and the ready availability of large datasets place multiple engineering fields at the threshold of dramatic progress. However, a unique challenge in NE is data scarcity because experimentation on nuclear systems is usually more expensive and time-consuming than most other disciplines. Particularly concerning is the lack of data for advanced reactor design and safety analysis, raising challenges for utilizing ML in licensing analyses of advanced nuclear reactors. In these cases, we need to move beyond “throw more data and re-train” at the problem, which is the common solution in areas such as computer vision and natural language processing that have access to “big data”. To this end, we are developing innovative data augmentation methods by leveraging deep generative learning and developing new techniques for nuclear-specific applications. Deep generative learning is an unsupervised ML technique that aims at discovering and learning the regularities or patterns in existing data in order to generate new samples that plausibly could have been drawn from the real dataset. In fact, a successful generative model may be able to generate new data that are indistinguishable from the real data. DNNs are usually used as the build blocks for deep generative learning models, where they are trained to approximate the sophisticated probability distributions underlying the real data. The general idea is to estimate and maximize the likelihood of the real data and create new synthetic data from the learned underlying distributions. In this way, the dataset can be significantly expanded, enabling us to train more accurate SciML models for other tasks and reduce over-fitting which is common when the dataset is limited.