We have finished some initial investigation on the data and operator neural networks. The progress page documents the full pipeline so far — feel free to review it before reading on. Below are five potential angles for where this project goes next. We will likely pick one or two to pursue seriously.
View Progress PageAfter training DeepONet on real indoor sensor measurements, the operator encodes some implicit physical law — but is it the one we would assume? My initial results suggest a clear sim-to-real gap exists: simply imposing a known PDE at the loss function is not enough, because the real environment does not perfectly obey that equation. The operator learns what the physics actually is here, not what textbooks say. The idea is to extract and characterise that learned physics, then test it on held-out data the model has never seen, and compare it directly against a black-box baseline under the same conditions. This matters because everyone right now is excited about PINNs, but almost all published results are on CFD or synthetic data. We would be doing this with real wireless sensor data from a real building, which is a fundamentally different and messier problem.
primary candidate novel contribution real sensor data physics-informedThere is a growing family of operator networks — DeepONet, FNO, GNO, and others — but almost all benchmarks are run on clean, simulation-generated data. No one, to my knowledge, has done a systematic comparison of these architectures on real indoor sensor data, where you have noise, irregular sensor placement, and genuine physical complexity. If the benchmark is rigorous and the dataset is well-documented, this could serve as a solid reference paper for anyone building surrogate models for built environments. Needs more literature search to confirm the gap, but the angle feels viable.
secondary candidate benchmark / reference FNO / GNO / DeepONetGiven a trained operator network, we can run an ablation over sensor subsets to find which nodes are essential for accurate temperature field reconstruction and which are redundant. This has practical value — if you are deploying a real building monitoring system, you want to know the minimum number of sensors and their best placement before you buy hardware. The analysis itself is fairly straightforward, and the result is immediately actionable. Honest assessment: not the most exciting research question, but it is useful and publishable.
applied / practical sensor placement ablation studyThe paper that published this dataset assumes temperature and humidity drive RSSI signal strength. My intuition is the reverse: RSSI is primarily a proxy for human activity (people moving, doors opening, devices being used), and it is that activity which then influences the thermal environment. If that is true, RSSI becomes a leading indicator rather than a dependent variable, which changes how you would use it in a model. This could be a clean causal analysis, but the impact is probably limited. Worth a section in a paper, maybe not a standalone contribution.
exploratory causal inference RSSIThe dataset includes two different labs with different geometries, furniture layouts, and occupancy patterns — and therefore different underlying physics. The question is whether a model trained entirely on Lab 1 can transfer to Lab 2 with no additional data (zero-shot), or whether a small number of Lab 2 observations is enough to adapt it (few-shot). If zero-shot transfer works at all, it would suggest the operator has genuinely learned something about indoor thermodynamics as a class of problem, not just memorised Lab 1. If it fails, the few-shot question becomes: how much data is the minimum needed before the model becomes useful? This is technically the hardest angle here, but it is also the one with the clearest scientific story and the broadest implications for deploying digital twins in buildings we have never instrumented before.
primary candidate high impact transfer learning zero-shot / few-shot cross-domain generalisation