A Phenomenological AI Foundation Model for Physical Signals

We explore the development of an AI foundation model that can be universally applied to physical processes of any nature. Our approach is based on a phenomenological framework, meaning that no prior physical knowledge or inductive bias is introduced. The aim is to construct a single, versatile AI foundation model capable of generalizing across diverse physical phenomena, domains, applications, and sensing apparatuses. This work is inspired by recent advancements in natural language processing (NLP), where generative models based on transformer architectures, such as GPT-4, have demonstrated that a single model trained on a vast corpus of text in self-supervised manner can perform as well as or better than specialized models across a range of tasks [8, 19]. In this paper, we present the design and evaluation of a physical AI foundation model, trained on 0.59 billion physical measurements covering a diverse range of real-world processes.