HYBRID PINN–LSTM ARCHITECTURE FOR PREDICTING THE DISPERSION OF CHEMICAL EMISSIONS IN THE ATMOSPHERE

Abstract

This paper proposes a hybrid approach based on a Physics-Informed Neural Network (PINN) that organically combines the fundamental convection-diffusion equations with recurrent LSTM layers. Such a combination makes it possible to effectively assimilate real-time data coming from a distributed network of IoT chemical sensors. Thanks to this, the model is capable of accounting not only for the physics of the pollutant transport process but also for the dynamics of changing meteorological conditions as new measurements arrive. The model was trained on 12,000 synthetic scenarios generated using high-fidelity CALPUFF models and Large Eddy Simulations (LES). This volume of data allowed us to cover a wide range of meteorological conditions, emission types, and terrain configurations. Furthermore, the architecture is equipped with a Bayesian uncertainty module that estimates prediction confidence intervals — a critically important feature for decision-making under emergency release conditions. Validation of the model was performed on real data from an industrial release that occurred in the Rivne region in 2023. The results confirmed the high efficiency of the proposed approach: the root mean square error (RMSE) for predicting ammonia concentration at distances up to 5 km from the source was 0.18 mg/m³. This is twice as good as the performance of the baseline SLAB model, which is traditionally used for modeling dense gases. Thus, the developed hybrid PINN–LSTM architecture demonstrates significant potential for the operational forecasting of hazard zones during chemical accidents under conditions of incomplete or noisy input data.