Sandia Labs FY21 LDRD Annual Report


Risk-adaptive experimental design for high-consequence systems. Credible assessment of high-consequence systems requires the fusion of simulation and experimental data. Currently, the nuclear weapons qualification process is data-starved and based on experimentation

and engineering judgment with some simulation guidance. Each aspect is conducted independently without statistical mapping, leading to exaggerated margins and over-designed components. This LDRD team developed an automated and goal-oriented framework to facilitate the rapid online design of experiments that target conservative measures of prediction uncertainty. Their approach minimizes the tail average of the prediction variance to produce certifiable and realistic margins that are robust to data corruption and noise. Researchers also developed specialized algorithms to solve such design problems, yielding a 200-fold reduction in computational effort . (PI: Drew Kouri)

Optimally placed accelerometer locations (green arrows) on a BARC ground- based test as determined using risk-adapted experimental design.

Adapting secure multiparty computation to support machine learning in radio frequency sensor networks.

This LDRD project team developed theoretical and practical foundations for secure and intelligent decentralized networks of low-power sensors that communicate via radio frequency. These networks are resilient to the random failure of a small fraction of nodes, remain secure even if an adversary captures a small subset of nodes, and are capable of basic machine learning. This new privacy-preserving machine- learning capability will help address national security priorities such as physical security and nuclear command, control, and communications. (PI: Jonathan Berry)

Privacy-preserving machine learning is demonstrated on a large network of autonomous drones at the AutonomyNM Robotics Lab.



Made with FlippingBook Ebook Creator