This Expedition in Computing established the Computational Sustainability Network (CompSustNet), which focused on the emerging field of Computational Sustainability (CompSust). Computational Sustainability harnesses the power of Artificial Intelligence (AI) and computation to forge innovative solutions addressing environmental, societal, and economic challenges for a sustainable future. CompSustNet expanded the research and community-building expertise of the Cornell Institute for Computational Sustainability (ICS), which successfully launched the new field. CompSustNet's core mission was to bridge connections between sustainability and computing and information science, both nationally and internationally, laying the foundation for computational sustainability to evolve into a self-sustaining field. CompSustNet’s research institutions included Cornell, Oregon State, Princeton, Vanderbilt, Bowdoin, Caltech, CMU, GeorgiaTech, Harvard, Howard, Ohio State, Stanford, UMass-Amherst, and USC.

One CompSustNet sustainability research theme was balancing environmental and socioeconomic needs. For poverty mapping, we used transfer and other machine learning approaches to estimate various large-scale spatial and temporal socioeconomic indicators from publicly available satellite and remote sensing data, culminating in SustainBench, a suite of benchmarks and datasets related to reducing hunger and poverty, and other United Nations Sustainable Development Goals. We developed influence maximization algorithms for peer-led HIV prevention, illustrating how AI algorithms can have real impact on homeless youth. In epidemiology, we studied evolution and prevention of infectious diseases in fisheries, modeled anthrax spread from animals to humans, and disease infections, including cholera. We developed crowdsourcing approaches to rebalancing Citi Bike's fleet, combining optimization and behavioral economics to design incentive mechanisms, more effective than staff transporting bikes with trucks.

Our second theme was biodiversity and conservation. We developed species distribution models to estimate species diversity and composition at continental scales and, more generally, multi-label and multi-target regression models, combining deep learning and reasoning. We used cameras and genetic monitoring to identify locations visited by animals to design wildlife conservation corridors that better connect existing habitat reserves. Combining home ranges and how animals use the landscape, we developed new ways to choose best habitats to conserve and protect multiple species with restricted budgets. To combat poaching, we used green security games to design randomized patrol strategies that work against sophisticated and knowledgeable poachers. In ecology, we studied system dynamics with long transients (temporary changes), how they relate to more permanent changes, and the resilience of systems subjected to regularly repeated disturbances.

Our third theme was renewable and sustainable energy and materials. We developed exact and approximate (with optimality guarantees) approaches to compute Pareto frontiers for strategic planning of hydropower expansion in the Amazon basin. By analyzing ecosystem tradeoffs for different energy production targets, we show foregone benefits to ecosystem services and energy production due to the lack of coordinated hydropower planning. Moreover, even though hydropower is often associated with clean energy, we find that greenhouse gas emission intensities for some suboptimal dam portfolios may be worse than burning conventional fossil fuels. In contrast, optimal portfolios can be as low as solar power. In search for new energy materials, including solar fuels, we developed the Deep Reasoning Networks framework (DRNets), which pushes the frontiers of AI for scientific discovery by seamlessly integrating reasoning about prior scientific knowledge into deep learning using an interpretable latent space. As a result, DRNets require only modest amounts of (unlabeled) data, in sharp contrast to standard deep learning approaches. DRNets reaches super-human performance for crystal-structure phase mapping, a core, long-standing challenge in materials science, enabling the discovery of fuel cells and solar fuel materials. We further showed the generality of DRNets for other tasks, including demixing overlapping Sudokus, and developed new ways to predict material properties. We developed new ways to combine optimization and machine learning to provide robustness guarantees for power systems like microgrids. Looking at consumer energy usage, we investigated patterns of interactions between household members around energy-related decisions and their role influencing home energy consumption.

Through interdisciplinary research projects (IRPs) based on our sustainability research themes, we identified and developed many cross-cutting computational approaches. The scale, scope, and complexity of the problems led us to advance the state-of-the-art in AI, machine learning, and other areas of computer science. These approaches focused on our three broad computational thrusts: constraint reasoning and optimization, dynamical models and simulation, big data and machine learning, multi-agent systems, crowdsourcing, and citizen science.

In addition to identifying and solving computational problems, we nurtured a vibrant Computational Sustainability research community. We expanded the CompSust conference series into ACM's COMPASS conferences and annual doctoral consortia hosted by different member universities. We organized relevant special tracks, workshops, symposia, master classes, tutorials at many AI conferences, and seminar series featuring senior researchers and graduate students. We gave numerous keynote speeches and public lectures related to computational sustainability. We trained PhD students and postdocs who obtained tenure-track positions at leading institutions. More information about these activities is on our website (https://www.compsust.net/) and connected social media.

Last Modified: 03/22/2024
Modified by: Carla P Gomes