Synthetic Data Generation For Inference In Remote Sensing Of Nuclear Non-proliferation Activities

We aim to find methodologies for remote sensing of non-proliferation activities from multiple sensors at a nuclear facility. To compensate for the difficulty of collecting real-world data with which to train and evaluate machine learning models, we leverage extensive subject matter expertise to develop a synthetic dataset covering a range of system designs and operational parameters. We pose two inference questions: (1) distinguish among several hypotheses of testbed activity, and (2) determine if and when the reactor begins operation. We compare the capabilities and performance of several machine learning techniques trained on this synthetic dataset. We achieve strong performance on both inference questions, evaluated on a hold-out test set of the synthetic data. Positive results are also obtained on the limited real data we have available.