The rapid advancement of foundation models offers a transformative opportunity to address some of the most pressing challenges in sustainable development. Unlike traditional machine learning models that are typically designed for single-task solutions, foundation models--especially Large Language Models (LLMs), Vision-Language Models, Multimodal Foundation Models, and Time Series Foundation Models (TSFMs)--are large-scale models trained on vast and diverse datasets. They are designed to handle multiple downstream tasks and offer high generalizability and adaptability.
In recent years, there has been growing interest in developing domain-specific foundation models to address the challenges of building and deploying these models in specialized contexts. Within the SIGEnergy community, there is growing momentum to develop, analyze, and explore the capabilities and limitations of such models, and to assess their adaptability across a range of tasks. The International Workshop on Foundation Models for Energy-Efficient Buildings, Cities, Transportation, and Sustainability (FMSust) provides a timely platform for researchers and industry practitioners to exchange ideas and share their latest findings, with the goal of advancing our collective understanding and responsible use of foundation models in the energy and sustainability domains.
Foundation Models for Energy and Sustainability
Datasets, Benchmarking, and Evaluation
Deployment, Validation, and Impact Assessment
Cross-Cutting Themes
*All times are local (Banff, Alberta, Canada)(UTC−6)
One Model to Run Them All? Foundation Models in the Physical World
Mario Berges (Carnegie Mellon University)
Foundation models (FMs) are poised to revolutionize the way we design, develop, and deploy cyber-physical systems (CPS). Their promise is a world in which each cyber-physical system can rely on a single model to complete arbitrary tasks via its sensors and actuators, paired with a natural-language interface for specifying and executing those tasks on the fly. This is a clear paradigmatic shift that FMs could make possible. However, the promised land remains elusive in the physical domains, especially those that matter the most for sustainability, such as buildings, energy and physical infrastructure. These CPSs are heterogeneous, data-scarce, and unforgiving, and not every task they carry out is even a good candidate for an FM to begin with. In this talk I begin by describing a vision for the role of foundation models in CPS, one that recognizes their limitations in physical domains. I then offer a few examples of research toward this vision from recent work in my lab, and close with some thoughts on where we, as a community, should go from here.
