Advancing Sustainability Through Data Science: Conference Highlights

On Friday, April 17th, over a hundred early-career data scientists, students, and professionals gathered at the Simonyi Conference Center in Stanford's Computing & Data Science (CoDa) Building for the 2026 Sustainability Data Science Conference—a dynamic day of sharing, discovery, and inspiration for those working at the intersection of data science and sustainability.

With opening remarks from David Lobell, Benjamin M. Page Professor in the Department of Earth System Science and the Gloria and Richard Kushel Director of the Center on Food Security and the Environment, the day started with an emphasis on the significance of the interdisciplinary work being done at Stanford right now, merging cutting-edge methods in AI and data science with domain-specific collaboration to make far-reaching impacts. Over coffee, lunch, and 19 total presentations across 3 sessions, attendees shared their insights with each other on pressing sustainability problems ranging from infill architecture for urban communities to battery health estimation in electric vehicles.

We wish to highlight some of our presenters who not only shared innovative methods and exciting discoveries, but also delivered engaging presentations that left an impact on us. 

Congratulations to Our Best Presentation Award Winners!

• 🏅 Ekaterina Landgren – Best Postdoc Presentation – Climate attribution in local news coverage of natural hazards
• 🥈 Joseph Janssen – Best Student Presentation – 2nd place – On the Adversarial Robustness of Hydrological Models
• 🏅 Aldis Elfarsdottir – Best PhD Presentation – Corporate Carbon Accountability
• 🥈 Mofan Zhang – Best PhD Presentation –  2nd place – Adaptive energy planning unlocks diverse decarbonization pathways and mitigates lock in risks
• 🌟 Jason Hu – Undergraduate Rising Star Award – Probing the Mechanistic Understanding of an AI Foundation Model Fine-tuned for Air Quality Forecasting
• 🌟 Yuchen Li – Undergraduate Rising Star Award

Best Postdoc Presentation: Ekaterina Landgren - Climate attribution in local news coverage of natural hazards

As action on climate change issues in the United States continues to critically hinge on public and political support and impact of natural hazards related to climate issues continues to grow, the information environment produced by media coverage of climate in regards to natural hazards becomes critical. In her presentation, Ekaterina Landgren, a postdoctoral fellow at the Stanford Doerr School of Sustainability, sought to answer 3 key questions:

• Does local news coverage reflect what scientists know about climate attribution?
• How do newspapers frame climate-related connections for all these different hazards?
• Do local demographic characteristics predict some variation in the frequency of these climate hazard connections in the news?

From a corpus of 75,000 hazard-related articles from local newspapers from 2,000 U.S. counties in the last 3 decades, Landgren used natural language processing (NLP) to investigate. 

First, Landgren finds that flooding and hurricanes are comparatively linked to climate change less frequently than drought and wildfire, despite potentially stronger attribution confidence. 

Drought, wildfire, and heat waves are most frequently causally connected to climate change, but coverage is primarily contextual in nature over causal - meaning climate change is mentioned in the same article as the natural hazard, but not cited as the explicit cause. Finally, Landgren finds that Democratic-leaning counties are more likely to link natural disasters to climate change. This research sheds light on how media coverage may reinforce certain preconceived notions about climate change, such as linkages to hot and dry hazards over wet hazards that are equally due to climate change, and play into partisan biases. What is certain though, is the need for scientists to continue to think critically about accurate, responsible, and effective communication to spur action on climate issues.

Best Postdoc Presentation - 2nd place: Joseph Janssen - On the Adversarial Robustness of Hydrological Models

Research in the field of computer vision and machine learning has shown that deep models can be susceptible to adversarial noise that maintains the appearance of the original data, but perturbs it in such a way that the model interprets it completely incorrectly. As a canonical example, a cat with adversarial noise added still appears as a cat to the human eye but becomes classified as an airplane by an object detection model. In his research, Janssen, a postdoctoral scholar in geophysics, examines this issue for machine learning based hydrological models that use precipitation, temperature, and evaporation, to predict stream flow for a geographic area. 

Using the adversarial fast gradient sign method, Janssen perturbed an Long Short-Term Memory (LSTM) deep learning streamflow prediction model along with a more traditional simulation-based Hydrologiska Byråns Vattenbalansavdelning (HBV) hydrology model. In a perhaps unexpected turn of events, Janssen found that the LSTM deep learning model was actually more robust to adversarial perturbation than the HBV model, as measured by a multitude of different metrics. With streamflow models in hydrology being key to flood preparation, water resource management, and engineering infrastructure safety, Janssen’s work paves the way towards trust and reliability in advanced new hydrology technologies.

Best PhD Presentation: Aldis Elfarsdottir -  Corporate Carbon Accountability

In her research, Aldis Elfarsdottir (PhD student in Management Science and Engineering) examined a critical question in corporate climate action: can we tell whether firms’ emissions reduction commitments are actually credible? Drawing on a decade of global survey data, Elfarsdottir combined firm-level disclosures with advanced natural language processing, specifically language models fine-tuned on climate coverage, to analyze how companies communicate their climate strategies. By constructing novel textual features such as specificity, climate-related language intensity, and audience targeting, and linking these to future emissions outcomes, the study moved beyond simply evaluating stated targets to understanding the signals embedded in how those targets are described.

The findings reveal that not all climate commitments are equally meaningful. While prior work often treated the presence of emissions targets as a positive signal, this research shows that targets alone can correlate with higher future emissions unless paired with detailed, specific plans. Firms that provide concrete, actionable disclosures, such as clearly defined initiatives and consistent reporting histories, are significantly more likely to achieve lower emissions over time. In contrast, heavy use of broad, aspirational climate language without substantive detail is associated with worse outcomes, suggesting potential greenwashing, particularly when firms tailor such messaging to specific audiences like investors or supply chain partners. Overall, the work highlights the importance of scrutinizing not just what companies promise, but how they communicate those promises, offering a more rigorous framework for evaluating credibility in the transition to a net-zero economy.

Best PhD Presentation - 2nd place: Mofan Zhang - Adaptive energy planning unlocks diverse decarbonization pathways and mitigates lock in risks

In her presentation, Mofan Zhang (PhD student in Civil and Environmental Engineering) tackled a central challenge in climate policy: how to make robust investment decisions for energy transitions under deep uncertainty. Traditional planning approaches rely on forecasting future technology costs and committing early to a single “optimal” pathway, but this strategy can lead to catastrophic lock-in costs if expectations prove wrong, especially for emerging technologies whose costs evolve dynamically through deployment like small modular reactors or advanced hydrogen. The research reframes the problem by emphasizing that future costs are endogenous: today’s investment decisions actively shape tomorrow’s technological landscape through learning-by-doing.

To address this, Zhang developed an adaptive, reinforcement learning–based framework for energy capacity expansion. Her approach continuously updates investment strategies in a closed loop, balancing near-term costs with the long-term value of information and technological learning. Across hundreds of simulated futures, the adaptive policy avoided the extreme cost overruns seen in static or myopic models by maintaining a diversified technology portfolio early on and pivoting decisively once clearer signals emerge. This strategy effectively hedges against downside risk while preserving upside potential, demonstrating that flexibility and strategic exploration can significantly reduce the risk of costly transition failures. The work highlights the importance of adaptive planning in guiding multi-trillion-dollar energy investments, offering a more resilient pathway toward achieving net-zero goals under uncertainty.

Undergraduate Rising Star Award: Jason Hu - Probing the Mechanistic Understanding of an AI Foundation Model Fine-tuned for Air Quality Forecasting

In this talk, Jason Hu (‘26, Undergraduate in Computer Science) explored a central challenge in atmospheric science: how to make fast, accurate forecasts of complex Earth system processes using AI foundation models while maintaining scientific reliability. The rise of data-driven foundation Earth AI models like Aurora can produce forecasts orders of magnitude greater in scale than before. However, these models lack built-in physical and chemical constraints, raising concerns about whether they truly capture underlying mechanisms. The research question, therefore, is not just how well these models perform, but how they work, and whether understanding their internal behavior can guide better model design. To address this, Hu applied a suite of interpretability techniques, including dimensionality reduction and sparse autoencoders, to probe the latent representations and internal structure of the Aurora model.

The findings reveal a nuanced picture. On one hand, Aurora demonstrates encouraging signs of physically consistent behavior, such as encoding meaningful global patterns and capturing phenomena like day-night cycles. On the other hand, it struggles significantly with extreme events, such as wildfire smoke plumes, where it exhibits “smearing” behavior that averages localized dynamics critical for real-world decision-making. By disentangling the model’s hidden features, the work also uncovers both meaningful atmospheric transport patterns and artifacts introduced by the model’s architecture. To support further analysis, Hu introduced AuroraScope, an open-source interpretability toolkit for systematically examining these internal mechanisms. Overall, Hu’s work highlights that improving AI-based Earth system models will require not just more data or compute, but deeper architectural insight. It offers an important step toward more trustworthy, high-fidelity environmental forecasting.

Congratulations again to our award winners, and thank you to the organizers of this conference- David Lobell, Sebastian Heilpern, Ching-Yao Lai, Laura Ventura, Haojie Wang, Veda Sunkara and Siddharth Sachdeva for their hard work in making this conference possible. Keep up the great work, and hope to see you next year!