Doctoral students across leading universities are discovering that artificial intelligence and spatial data science can illuminate some of society's most pressing challenges, from healthcare inequities to clean energy distribution. Rather than building the next large language model, these researchers are combining machine learning with geospatial mapping to answer real questions about who gets access to treatment, where renewable energy infrastructure is built, and how historical infrastructure shaped modern cities.

What Are PhD Students Actually Building With AI and Spatial Data?

At Rutgers School of Social Work, doctoral students enrolled in a spring course called "AI & Spatial Data Science for Global Development" tackled five distinct research projects that showcase how computational methods can address social welfare challenges. The course, taught by Assistant Professor Dr. Woojin Jung, brought together students from social work, public policy, computer science, economics, and geography to work collaboratively on policy-relevant research.

The projects spanned critical areas of public concern. One team examined geographic disparities in substance use treatment access for Medicare beneficiaries across New Jersey, identifying underserved areas where older adults struggle to find care. Another team investigated how the introduction of electric streetcars in late-nineteenth-century New York shaped neighborhood demographics and economic outcomes by combining census data with historical maps. A third project mapped the spatial mismatch between renewable energy production sites and electric vehicle charging infrastructure across New York State, revealing that clean energy facilities are concentrated in lower-income rural areas while charging stations cluster in wealthier urban centers. A fourth team analyzed how geography and digital infrastructure predict AI engagement in the Global South, suggesting that location-based factors may be stronger predictors than traditional demographic variables.

"Taking Professor Jung's well-structured course has been an incredibly rewarding experience. The lectures perfectly introduced us to the applications of spatial analysis in social sciences, while the lab sessions led by the two TAs gave us great hands-on practice with AI analysis methods," said Huiyuan Zheng and Xu Zhang, researchers who studied energy equity in New York State.

Huiyuan Zheng and Xu Zhang, PhD Students at Rutgers School of Social Work

How Are These Students Combining Machine Learning With Geographic Data?

• Satellite and Crowdsourced Data: Students worked with innovative data sources including satellite imagery, crowd-sourced maps, and social media data to conduct granular needs assessments and support evidence-informed policy design.
• Python Programming and Geospatial Tools: The course provided hands-on training in Python programming, with supplementary use of QGIS, a specialized software for geospatial analysis that allows researchers to visualize and analyze location-based data.
• Historical Data Integration: Teams combined contemporary datasets with historical records, such as digitizing 1880 and 1899 maps of New York's streetcar networks to trace long-term patterns of urban transformation using difference-in-differences statistical frameworks.

The interdisciplinary approach proved essential. One team studying substance use treatment access noted that "each of us within this group brought different skills and experience, since the class drew students from different disciplines, programs, and years across the university, which allowed us to learn from one another". This cross-disciplinary collaboration mirrors a broader trend in AI research where domain expertise in social work, policy, or environmental science becomes as valuable as machine learning skills.

What Did the Energy Equity Research Reveal?

The findings from the New York State energy infrastructure study highlight a critical fairness question often overlooked in climate policy discussions. After analyzing more than 5,000 census tracts across the state, researchers confirmed a stark spatial mismatch: solar and wind facilities are concentrated in lower-income rural areas due to land availability and lower property costs, while electric vehicle charging infrastructure clusters in wealthier urban cores driven by income and market demand.

"This means some communities bear the landscape costs of clean energy production while others enjoy most of the consumption benefits," the research team explained. "Without deliberate policy intervention, the green transition risks reinforcing old inequalities". This finding demonstrates how AI-powered geospatial analysis can expose hidden patterns in policy outcomes that might otherwise remain invisible to policymakers.

How Are Leading Research Labs Advancing Machine Learning for Science?

Beyond social applications, cutting-edge research labs are pushing machine learning into scientific domains where accuracy and interpretability matter enormously. The Dynamic Systems Lab (DSL) at University College London presented two major advances at the International Conference on Machine Learning (ICML) 2026 in Seoul, South Korea, in July.

The first breakthrough, called "MMPD-Bench," introduces the first large-scale benchmark for AI-driven Mueller matrix polar decomposition, a traditionally physics-based inverse problem in computational polarimetry and scientific imaging. The benchmark provides over 21,000 real-world polarimetric measurements and specialized evaluation datasets, establishing a new foundation for developing trustworthy AI surrogates in scientific imaging. This work reframes a classical physics problem as a machine learning task, potentially accelerating scientific discovery in optics and materials science.

The second contribution, called "Augmented Lagrangian-Guided Diffusion" (ALGD), combines diffusion-based policy generation with constrained reinforcement learning to create safer decision-making systems. By reformulating the Lagrangian as an energy function and introducing an augmented optimization strategy, ALGD stabilizes policy learning while preserving both safety and optimality. The framework demonstrates how generative AI can move beyond offline policy learning toward reliable and safe decision-making in complex control systems, with potential applications in robotics and autonomous systems.

"The presentation attracted interest from researchers working at the intersection of machine learning, computational imaging, and scientific AI. Discussions focused on the growing role of physics-aware representation learning, multimodal generation, and neural operators for accelerating scientific discovery while preserving physical consistency," noted the DSL team.

Dynamic Systems Lab, University College London

Why Does the Gap Between AI Predictions and Real-World Drug Binding Still Matter?

In pharmaceutical research, machine learning has scaled drug-target interaction prediction from thousands of compound-protein pairs to billions, yet a persistent challenge remains: the gap between computational predictions and confirmed laboratory results. Deep learning models trained on chemogenomics data now rank drug candidates across entire proteomes in hours rather than months, but false positives still consume significant experimental resources.

Chemogenomics treats chemical space and biological target space as a single, linked system rather than two separate problems. This approach maps compounds against families of related proteins, such as kinases or G protein-coupled receptors, so that binding data collected for one target informs predictions for related targets. The strategy enables transfer learning, where a model trained on kinase inhibitor data can apply some of that learned chemistry to a related, poorly characterized kinase.

Public benchmark datasets like BindingDB, which contains more than 2.9 million binding measurements across 1.3 million compounds according to a 2025 database update, give every drug-target interaction prediction model a common yardstick. However, each dataset encodes its own measurement biases. DAVIS and KIBA are the two benchmarks most commonly used for kinase-focused binding affinity prediction, but model rankings can shift depending on which benchmark is used, since each dataset weights different regions of chemical and target space. This fragmentation highlights why uncertainty-aware modeling and tighter integration with structural and functional assays remain active research frontiers.

The common thread across these research efforts, whether at Rutgers, UCL, or in computational chemistry, is that AI's real value emerges when it is grounded in domain knowledge, rigorous benchmarking, and a clear understanding of where predictions diverge from ground truth. Students and researchers are learning that building trustworthy AI systems requires not just better algorithms, but deeper collaboration across disciplines and a commitment to transparency about what machine learning can and cannot reliably predict.