Sandia Labs FY22 Laboratory Directed Research & Development Annual Report

ADVANCING STRUCTURAL DYNAMICS CAPABILITIES WITH NONLINEAR MODAL ANALYSIS. Linear modal analysis has served as the

recent articles in Mechanical Systems and Signal Processing and IOPScience , and numerous conferences and society meeting presentations. The LDRD team collaborated with Sandia Alliance partners from the University of New Mexico, Georgia Tech, and Texas A&M University. They also collaborated with the Sandia National/Regional partners from New Mexico State University, ETH Zurich, Los Alamos National Laboratory, and ATA Engineering. (PI: Robert Kuether)

Damage caused by malware to applications in the national security sector is increasing at exponential rates. Leveraging ML techniques for its detection can achieve greater than 99% classification accuracy on benchmark datasets, but performance drops significantly in deployed settings The Sandia MALGEN LDRD team, in collaboration with Sandia National/Regional partner New Mexico Tech, explored and addressed this gap by proposing a novel evaluation method that held out a malware family from training and determined this type of evaluation provided more realistic results. Malware was detected by discerning commonly exhibited cornerstone of computational and experimental structural dynamics for many years. Nonlinearity, such as frictional contact or nonlinear materials, may invalidate the assumptions necessary to model the dynamics with linear system theory. The Sandia LDRD team set out to extend the computational and experimental modal analysis framework to characterize, identify, and predict vibration response of nonlinear mechanical systems. The novel “nonlinear phase resonant mode” theory provides a consistent mathematical definition for both simulation and test methods. The nonlinear modes are identified experimentally using active controllers to maintain a phase quadrature condition. The simulation counterpart uses the multi-harmonic balance method with continuation to efficiently compute the periodic orbits from nonlinear models. The generalized, nonlinear modal analysis framework is applicable to arbitrary conservative and non-conservative nonlinearity. This project resulted in ten publications including

Nonlinear phase resonant mode testing shows: (a) a Phase resonance test of a nonlinear beam assembly, and (b) a correlation of measured and predicted amplitude dependent natural frequency and damping ratio.

MALGEN: MALWARE GENERATION WITH SPECIFIC BEHAVIORS TO IMPROVE MACHINE LEARNING-BASED DETECTORS.

behaviors such as Dynamic Link Library injection and process hollowing, which goes beyond signature based methods by (1) assuming a set of behaviors common to most malware samples exist, and (2) the detection of behaviors will generalize to novel malware. Performance was further improved by generating additional synthetic malware with specific behaviors. The methods determined through this three-year project lay a foundation for improved integration of ML that have been highlighted in two journals and four conferences/workshops. (PI: Michael Reed Smith)

71

LABORATORY DIRECTED RESEARCH & DEVELOPMENT