Summary
Highlights
Professor Carla Gomes introduces the field of computational sustainability, which leverages AI and computational methods to address challenges in sustainable development. She highlights the broad scope of sustainable development, encompassing environmental, economic, and societal goals, as defined by the UN's 17 Sustainable Development Goals. Gomes advocates for 'knowledge-centric AI' over purely data-driven approaches for scientific discovery and decision-making, emphasizing the need for systems that can reason from first principles, generalize beyond training data, be interpretable, and handle multi-objective optimization.
Gomes discusses AI's role in combating biodiversity decline through joint species distribution models, using bird observations from Cornell's eBird program. These models integrate citizen science data with environmental factors to predict species habitats and migrations at fine resolutions. She explains how these models inform conservation efforts, such as the 'Bird Returns' program, and details the computational challenge of modeling species interactions, which requires estimating a multivariate Gaussian model with an interpretable latent space.
The talk shifts to AI's application in material science, specifically crystal structure phase mapping. Gomes presents a 'deep reasoning net' framework that combines deep learning with knowledge-centric reasoning to infer crystal structures from X-ray diffraction patterns. She illustrates the concept with an analogy to solving overlapping Sudoku puzzles, demonstrating how physical rules and prior knowledge are incorporated into the model's interpretable latent space and differentiable decoder (Bragg's law) to demix complex material structures and discover new materials like solar fuels.
Gomes addresses the problem of strategic planning for hydropower dams in the Amazon basin, highlighting the negative environmental impacts despite their clean energy perception. She proposes using multi-objective decision-making and Pareto optimization to evaluate tradeoffs between energy generation and ecological value. By exploiting the tree-like structure of rivers, her team developed a fully polynomial-time approximation scheme (FPAS) using dynamic programming to efficiently compute Pareto frontiers, allowing for better-informed policy decisions in complex, multi-criteria scenarios.
The final topic covers the 30x30 initiative to protect 30% of land and water by 2030, focusing on freshwater fish conservation. Gomes introduces a novel NAPSAC variant to optimize protected areas, emphasizing how exploiting the tree structure of river networks allows for efficient solutions. She concludes by summarizing the broader vision of knowledge-centric AI, emphasizing its transferability across domains, the development of multi-agent frameworks like 'robot scientists,' and the reciprocal benefits of applying computational science to real-world problems.