ChiragManchanda

Next-generation environmental sensing systems

There is an awkward mismatch at the centre of air quality science. The spatial and temporal scales that matter most for health and policy are the ones we observe least. Fixed monitors, mobile platforms, aircraft and satellites all look at the same city, and each one sees a narrow, biased slice of it.

For my dissertation I built an assimilation system that combined mobile monitoring from Google Street View cars with fixed-site sensor networks, producing pollutant concentration fields at 30 meters and 15 minutes. It rests on compressive sensing rather than a prescribed inventory, so the spatiotemporal structure comes out of the data instead of being assumed going in. The maps showed sharp gradients along freight corridors and around small industrial clusters, features that conventional monitoring networks and chemical transport models both tend to smear away. They also turned out to be useful for decisions people were already trying to make: where to put the next monitor, which freight routes to target, how to draw a low-emission zone that lowers exposure on the blocks you meant to help.

Toward an urban pollution digital twin

A concentration map tells you where pollution ends up. It does not tell you who put it there. High-resolution maps make good diagnostics, but policy needs the attribution step.

I extended Bayesian inversion down to roughly 150 meters, fine enough to attribute emissions to a particular corridor, facility or neighborhood. What comes out is an emission estimate grounded in observations rather than in a bottom-up inventory. Couple that with a transport model and you can test an intervention before committing to it: rerouting freight, modernizing port operations, tightening controls at one specific facility. You also see how each option redistributes exposure, which matters, because the aggregate improvement and the distributional outcome are frequently not the same story.

The obstacle is cost. Inverse modeling leans on chemical transport models that need repeated case-specific simulations, which rules out large ensembles, real-time use, or tight coupling with an optimizer. Reduced-complexity models and unconstrained machine learning surrogates are cheaper but tend to give up physical interpretability and robustness across regimes.

Reverse-engineering environmental policy

Environmental policy is usually designed forwards. Propose a strategy, model the outcome, adjust, repeat. I am interested in running the loop the other way: start from the air quality and climate target, then solve for the ways to reach it.

During my Ph.D. I adapted inverse modeling to search across combinations of emissions reductions, spanning space, sector and chemical precursor, for the ones that meet air quality and climate goals together. Writing equity in as a constraint, instead of scoring it after the plan is drafted, changes which plans win. In California the model favored a set of targeted low-emission zones over broad regional cuts, and got more exposure reduction and less disparity out of them by exploiting how transport and chemistry interact. The limitation is that these models hold the world still. They cannot see how social and economic systems adapt once a policy is actually enacted.