Air pollution is a leading global health threat, yet many developing countries lack the dense monitoring infrastructure needed for accurate exposure assessment and informed policy. India, with a population of over 1.4 billion, operates only about 600 official PM₂.₅ monitoring stations. This data gap poses a fundamental barrier to protecting public health.
Deploying new sensors is costly, and deciding where to place them for the most valuable data is a massive computational challenge. The most accurate, information-theoretic methods (such as Mutual Information) are well-founded but prohibitively slow for large domains, since their runtime scales with the number of candidate locations.
Now, researchers from the Indian Institute of Technology Gandhinagar (IITGN), in a paper accepted to the AAAI Conference on Artificial Intelligence 2026 (AI for Social Impact track), have developed a new framework that resolves this long-standing “accuracy-versus-scalability” trade-off.
What is the new framework?
The approach reframes the discrete sensor-placement task as a continuous, differentiable optimization problem. Instead of evaluating every grid point, the framework treats each sensor’s (latitude, longitude) as a trainable parameter and uses gradient descent—the same technique behind modern AI – to “walk” an entire batch of sensors to their jointly optimal positions. The breakthrough: its runtime is independent of grid size.
“We were stuck in a trade-off. Fast, simple methods gave poor placements, often clustering sensors or pushing them to map edges,” says Prof Nipun Batra, Associate Professor and co-author, Department of Computer Science and Engineering, IITGN.
“Our approach achieves the accuracy of information-theoretic methods at a speed that’s finally practical. We can now find the best spots for 100 sensors just as fast as for one.”
Why This Framework Matters
This new approach makes high-end, information-theoretic sensor placement feasible at a national scale for the first time.
For policymakers and environmental agencies, this means they can now design and expand monitoring networks with confidence, ensuring every costly sensor provides the maximum possible information.
“Our AI avoids redundant clustering and boundary bias,” explains Zeel B. Patel, PhD student and co-author, Department of Computer Science and Engineering, IITGN. “It learns to distribute sensors intelligently, improving both accuracy and coverage.”
The results are striking. On a continental-scale dataset for India, the new method achieved a 4% greater reduction in prediction error (RMSE) compared to the widely-used “Maximum Variance” heuristic. In a focused regional study, it approached the accuracy of the “gold standard” Mutual Information method, while beingorders of magnitude faster.