Code, trained PINN models and derived results for: Air Purifier Performance in an Occupied Office: Field CADR and Effectiveness from a Reconstructed Device-Off Condition

Training code, analysis code, trained physics-informed neural network (PINN) weights for thirty paired CO2 and PM2.5 networks, fitted input scalers, the member list and per-member accuracy, derived per-state results, the synthetic parameter-recovery experiments for both networks, the networks of the structural and humidity sensitivity tests, and the underlying field measurements for the counterfactual evaluation of a portable air purifier in an occupied office. The deposit reproduces the reported air change rate, particle penetration factor, removal rate, field CADR and effectiveness results of the second revision of the manuscript (Building and Environment, under review, BAE-D-26-05957). Version 3 replaces the ensemble of the first revision with thirty network pairs with softplus outputs, all of which are used, corrects the room volume to 71.25 m3 and the temperature input of the PM2.5 networks, and adds synthetic PM2.5 recovery tests, a structured removal-rate variant and a humidity sensitivity test. The weights of the first-revision networks remain in version 2. See README.md for the layout, the archives and the scripts.