Sandia Labs FY21 LDRD Annual Report

FY21 ANNUAL REPORT

Developing engineered bioremediation technologies for breaking down waste. By applying gene editing tools and metabolically engineering microorganisms, LDRD researchers determined the key enzymological parameters needed to influence the kinematic activity of polymeric degradation via enzymatic activity for bioremediation. The team leveraged material characterization methods derived from dielectric permittivity changes within polymers to understand the material structure and composition changes necessary for developing engineered bioremediation technologies. The team created innovative capabilities (a) to characterize changes in complex permittivity and (b) provide materials information regarding changes in material composition and structure, such as degree of cross-linking, the amount of solvent uptake, and changes in porosity among others. Ultimately,

the research characterized kinetic reaction parameters for biodegradation, yielded understanding of metabolic by- products, and demonstrated the potential system scalability required for plastic waste bioremediation. (PI: Isaac Avina)

Biomaterial testbeds.

First step toward understanding device-aware probabilistic neural networks. Many critical applications that involve forecasting or predictions based in high uncertainty or complex scenarios could benefit from artificial intelligence (AI) capabilities that leverage probabilistic computing. Unfortunately, research into probabilistic neural networks has proven challenging due to the computationally prohibitive nature of producing large-scale random numbers. This LDRD project investigated how algorithms using probabilities to compute could be modified to take advantage of novel microelectronics devices that produce noise through their physics. Results provided preliminary evidence that probabilistically sampling neural networks with a paradigm consistent with potential noisy devices could lead to achieving uncertainty-aware AI. Researchers observed that sampling neural networks does not significantly impair performance and that uncertainty introduced by sampling appears to match intrinsic uncertainty in data. Furthermore, it appears that these algorithms may tolerate devices whose noisy behavior is of limited precision, an important consideration for translating this

approach to emerging hardware technologies. The research supports the team’s supposition that co- design between AI algorithms and emerging hardware technologies may yield advanced microelectronics capabilities. (PI: Brad Aimone)

Illustration of how sampling an artificial neural network can allow uncertainty in inputs to be appropriately be recognized by an AI algorithm.

46

LABORATORY DIRECTED RESEARCH & DEVELOPMENT