Vector AI x Science + Sustainability Workshop
at AI/CRV 2026

Measuring the biodiversity of our world is increasingly important, with a large number of species facing extinction in the coming decades. As part of the BIOSCAN project, scientists around the world are building an observation system for monitoring biodiversity. Advances in deep learning have the potential to greatly reduce the manual effort required in the workflow and data pipeline for collecting and cataloguing biological specimens. Moreover, we can enable a deeper understanding of the interaction between different species. In this talk, we will describe our efforts to leverage recent advances such as self-supervised techniques to learn improved representations of DNA barcodes, and contrastive learning to align DNA barcodes with images and other modalities representing biological species. The intersection of deep learning and biology in which our work is situated holds many challenges for future work from several research communities.

Angel Xuan Chang is an Associate Professor at Simon Fraser University and a Canada CIFAR AI Chair with Amii. Prior to this, she was a visiting research scientist at Facebook AI Research and a research scientist at Eloquent Labs working on dialogue. Dr. Chang’s research focuses on the intersection of natural language understanding, computer graphics, and AI. Her research connects language to 3D representations of shapes and scenes and addresses grounding of language for embodied agents in indoor environment, as well as 3D content creation from language. Her group also works on using machine learning for biodiversity monitoring, specifically with DNA barcodes as part of the larger BIOSCAN project. Her work has been recognized by awards such as the SGP dataset award (for ShapeNet and ScanNet).

Kelsey Doerksen is a joint postdoctoral fellow at Arizona State University in the School of Augmented Intelligence and Computing and the University of Cape Town's Climate Risk Lab, and a research affiliate at the NASA Jet Propulsion Laboratory. Her research focuses on combining her expertise in AI, Earth Observation, and Foundation Models to applied problems in the climate, food security, and biodiversity domains. She is a recent graduate from the Autonomous Intelligent Machines where she completed her thesis in the Oxford Applied and Theoretical Machine Learning Group on "Applied Machine Learning for Earth Systems Science", collaborating with the United Nations, European Space Agency and NASA during her PhD.

Society has always depended on Earth's resources, and the rapid growth of AI infrastructure, with its enormous demands for copper, energy, and water, is a timely reminder of the need to locate and responsibly manage such resources. Geophysics provides an essential means of non-invasively imaging the subsurface, and is itself being transformed by the very technologies that depend on it. At the heart of geophysical data analysis lies the inverse problem: given measurements at the surface, what can we infer about the subsurface? The approaches we take to solve the inverse problem, posing it as an optimization problem where we connect a forward model (usually a numerical simulation) to observations through iterative model updates, are not so different from training a neural network. Recognizing this parallel opens the door to hybrid approaches that combine the flexibility of machine learning with the physical constraints of physics-based inversion. I will present research from my group at UBC exploring this intersection, including hybrid ML and physics-driven inversion for mineral exploration and neural network approaches for detecting and classifying unexploded ordnance.

Lindsey Heagy is an Assistant Professor in the Department of Earth, Ocean and Atmospheric Sciences and Director of the Geophysical Inversion Facility at the University of British Columbia. Her research develops computational approaches to numerical simulation, inversion, and machine learning, with a focus on electromagnetic and potential field methods to characterize the subsurface from geophysical data. Applications of interest include mineral exploration, carbon sequestration, groundwater, and environmental studies. She is a co-founder of the SimPEG project, which develops open-source software for geophysics, used by researchers and practitioners worldwide.

Workflows that incorporate both AI techniques and scientific simulation hold the promise of opening new avenues of scientific discovery. Creating these new applications requires practitioners to bridge ideological gaps (e.g. first-principles vs. data-driven approaches) and technological challenges (e.g. embedding AI within C++ or Fortran code). This talk describes a number of research projects in computational fluid dynamics, weather/climate, and fusion modelling that demonstrate the value of these hybrid approaches. Lastly, we will discuss the abstract workflow concepts shared by these use cases and how we are incorporating the lessons we have learned into the open-source project RHAPSODY, co-developed by HPE and collaborators at Rutgers University and the Department of Energy. 

Andrew Shao is an HPC&AI Research Scientist within the AI Research Lab at Hewlett Packard Enterprise. He leads many scientific collaborations with industry, government, and academic partners exploring how to apply AI techniques in and around numerical simulations. He has a background in numerical model development having worked on the development of climate models in the US and Canada. He holds a PhD in Oceanography from the University of Washington and undergraduate degrees from the University of California, San Diego. When not staring at a screen, he can often be found volunteering at a cooperative farm near where he lives in Victoria, BC.

Markov Chain Monte Carlo (MCMC) is a powerful algorithmic framework for sampling from complex probability distributions. Standard MCMC methods struggle with high-dimensional distributions containing well-separated modes, becoming trapped in local regions. Parallel tempering (PT) addresses this by using intermediate annealing distributions to bridge a tractable reference (e.g., Gaussian) and an intractable target distribution. However, classical PT is inflexible, fragile, hard to tune, and prone to performance collapse on challenging inference tasks. This talk introduces non-reversible parallel tempering (NRPT), which provably dominates classical PT algorithms. We show that NRPT undergoes a sharp algorithmic phase transition with increased parallelism, becoming robust, easy to tune, and scaling efficiently to GPUs. I will then demonstrate how to further accelerate PT using neural transports such as normalising flows and diffusions. We demonstrate this framework across a variety of examples in Bayesian inference and inference-time control for diffusion models.

Saif  recently joined UBC an Assistant Professor in the Department of Statistics as part of the AI Methods for Scientific Impact cluster within CAIDA. Previously, he held a Florence Nightingale Bicentenary Fellowship at Oxford's Department of Statistics. His research focuses on the design, analysis and applications of annealing algorithms for scalable Bayesian inference and generative modelling, with applications spanning astronomy, chemistry, and biology. His work has been recognised with the Pierre Robillard Award from the Statistical Society of Canada, the Cecil Graham Award from the Canadian Applied and Industrial Mathematics Society, and an Honourable Mention for the Savage Award from the International Society for Bayesian Analysis.

Open-source AI is growing at an extraordinary pace. Model hubs now host millions of artifacts, and every foundation model can spawn hundreds or thousands of fine-tunes, adapters, quantizations, and forks. But here's the problem: we have almost no idea what any of this costs the planet. We argue that compute efficiency alone is not enough for sustainable AI. In fact, lower per-run costs can accelerate experimentation and deployment, potentially increasing the aggregate environmental footprint, not reducing it. And right now, the energy use, water consumption, and emissions across these derivative model lineages are rarely measured or disclosed in any consistent, comparable way. The ecosystem's true impact is essentially invisible. What we need is coordination infrastructure — something that tracks impacts not just for base models, but across entire model lineages. To that end, we propose Data and Impact Accounting (DIA): a lightweight, non-restrictive transparency layer for open-source AI. DIA does three things: it standardizes carbon-and-water reporting metadata; it integrates low-friction measurement into common training and inference pipelines; and it aggregates these reports via public dashboards to summarize cumulative impacts across releases and derivatives. The goal is to make derivative costs visible, support ecosystem-level accountability, and do it all without compromising openness.

Iulliia is a Research Manager and Postdoctoral Researcher at the University of Guelph and the Vector Institute, where she combines hands-on research with project management. After completing a PhD in Biomedical Engineering, her research evolved toward developing practical, scalable AI solutions for real-world challenges in biodiversity, conservation, and climate science. Alongside this work, she is interested in Sustainable AI, particularly the environmental footprint of AI systems and the development of more efficient, transparent, and responsible approaches to AI research and deployment. Her work sits at the intersection of AI, biology, and environmental science, with a strong emphasis on interdisciplinary collaboration and communicating research to both technical and broader audiences.

TBD