I-Corps: Machine Learning Algorithms and Tools for Analysis and Optimization of Infrastructure

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

The broader impact/commercial potential of this I-Corps project is to develop algorithmic methods and computational systems that will provide new information, cheaply, quickly, consistently, and at a large scale, to inform investments in building, energy, and green infrastructure for engineering consultancies, construction companies, power utilities, real-estate developers, and cities. Currently, studies for built, energy, and green infrastructure investments are typically made case-by-case and employ infrequent, manual, and expensive audits performed by trained professionals. In addition, the ground-truth data (site surveys and other urban asset databases) used to inform these decisions are siloed across stakeholders and are often expensive and difficult to collect. The ecosystem of simulation and analysis tools is fragmented into a variety of highly complex software packages that require considerable effort and customization. This results in much repeated effort, inconsistent results across different projects, and high costs for surveys and instrumentation. This I-Corps project will investigate commercial applications of a machine learning system that ingests, aggregates, processes and learns mappings between, remote-sensing, model simulation, and local sensor and survey data on buildings and their environment. The data used will be best-available open source and commercial satellite data, as well as Internet of Things data from buildings and energy infrastructure. The project further develops neural-network based generative machine learning models that can emulate complex physical and engineering systems but are much cheaper and faster to deploy and maintain in practice. This system would allow users to make inferences (physical parameter recovery, forecasts) about built, green, and energy infrastructure and their surroundings (micro-climates) where this data is not available, in a low-cost, large-scale, automated way, removing the need for certain expensive sensor deployments or manual surveys. The system develops a top-down, machine-learning based simulation capability that would allow to perform fast scenario and impact analysis of investments and interventions.