Modeling brake emissions using an ad hoc physics-inspired recurrent neural network

Brake emissions have gained increasing attention over the past twenty years. Still, due to the inherently transient and complex nature of brake emissions, advanced modeling techniques are necessary but remain limited in the existing literature. In this study, randomized tests were conducted on a reduced-scale dynamometer with braking parameters within the domain of the Worldwide Harmonized Light Vehicles Test Procedure (WLTP). Emissions were measured using an Optical Particle Sizer (0.3-10 μm) and modeled with a Neural Network (RNN) featuring an ad hoc recurrent architecture. The RNN model comprises: 1) an Internal State, s , which describes the bedding-in process and surface-state transitions caused by changes in braking parameters, and 2) a stationary component, e 0 , which depends solely on dissipated energy and brake deceleration. WLTP data were used to test the model. Additional randomized tests were then conducted, yielding strong R 2 values ranging from 0.96 to 0.99 for cumulative emissions and lower values of 0.55 to 0.60 for individual events. • Neural Network with ad hoc recurrent architecture was applied to emission data from dynamometer tests; • The steady state component is modelled as function of dissipated energy and brake deceleration; • The recurrent component can describe the bedding process and transition induced by varying brake parameters; • Introduction of new concept of steady state and emission-related states; • High determination coefficient (> 0.95) for cumulative emissions and intermediate for individual brake events (0.55-0.6) on data from WLTP and randomized tests.