SBIR Phase I: Synthesizing large and diverse data-sets for training machine learning algorithms using physical modeling and simulation

The broader/commercial impact of this Small Business Innovation Research (SBIR) Phase I project, would be to diminish the hurdles in building a synthetic data generator for robust automated detection technologies. Passenger and personal property screening is an essential component of Department of Homeland Security’s (DHS) strategy to combat terrorism and targeted violence. The synthetic data generated by the proposed technology can be used to train as well as characterize the advanced screening solutions deployed. This will boost understanding of expected field performance and hence confidence in systems used to protect people and critical infrastructure. At airports, fewer false alarms from people and baggage screening equipment would translate to shorter lines, smaller wait times and decreased stress levels. Better threat detection rates would boost confidence in the screening solutions and truly help in reducing anxiety surrounding air travel, large gatherings, and outdoor events. The apparatus for generating synthetic data can also benefit education and training of the budding STEM workforce in advanced technologies. This Small Business Innovation Research (SBIR) Phase I project aims to demonstrate that a novel radiation physics solver based on first principles can generate synthetic data that matches the realism of data obtained through manual acquisition on physical radiation-based scanners. In addition, the solver can grow/widen the sample probability distribution to more closely match the population probability distribution than manually acquired data and do so in a hitherto unrealized linear computational time. In an emerging world of machine learning based automated threat or anomaly detection, this cost-effective data synthesis fulfills an immediate need to address the problem of data paucity to both train and test such algorithms. The research and development effort in this Phase-I project will be focused on developing computational methods to estimate the residual energy post photon-matter interaction in a cost-efficient manner. Representative object assemblies will be constructed to virtually scan and generate realistic and precisely annotated imaging data. Appropriate metrics will also be developed to measure the quality of the created synthetic images. The challenge will be to match the resolving power of the relevant modality as it applies to specific application areas. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.