AI Safety for Physical Infrastructures: A Collaborative and Interdisciplinary Approach

Abstract

Where AI systems are increasingly and rapidly impacting engineering, science, and our daily lives, progress in AI safety for physical infrastructures is lagging. Most of the research and educational programs on AI safety do not consider that, in today's connected world, safety and security in physical infrastructures are increasingly entangled. This technical note sheds light, for the first time, on how computer science and engineering communities, for example, mechanical and civil, can collaborate on addressing AI safety issues in the physical infrastructures and the mutual benefits of this collaboration. We offer examples of how probabilistic views of engineers on safety can contribute to quantifying critical parameters such as “threshold” and “safety buffer” in the AI safety models, developed by the world-leading computer scientists. We also offer examples of how novel AI and machine learning tools, for example, do-operator, a mathematical operator for intervention (vs. conditioning); do-calculus, machinery of causal calculus; and physics-informed neural networks with a small number of samples can help fatigue and fracture research. We envision AI safety as a process, not an object, and contribute to realizing this vision by initiating a collaborative and interdisciplinary approach in establishing this process.