A Latent Space Framework for Modeling Transient Engine Emissions Using Joint Embedding Predictive Architectures

Accurately modeling exhaust emissions during transient events, such as rapid acceleration and deceleration, remains a long-standing challenge in the automotive field and is a prerequisite for real-time control and optimization in modern powertrain systems. The nonlinear emission behavior is difficult to capture, and early mathematical models frequently relied on approximations that limited their accuracy. With the rise of data-driven methods, especially deep neural networks, including multilayer perceptrons and long short-term memory (LSTM) networks, modeling capabilities have improved significantly. However, to achieve meaningful accuracy, these models often require deep architectures, which increase computational load and inference time. This makes their application in real-time control systems challenging, as controllers must run at high sampling rates while maintaining accuracy, which is often a contradictory requirement. This paper addresses the trade-off with a new framework built on the concept of latent space-encoded modeling. Using data collected under real driving conditions from a Portable Emission Measurement System (sampled at 1 Hz) and complemented by Hardware-in-the-Loop test bench data (5 Hz), the use of Joint Embedding Predictive Architectures (JEPA) is explored to derive a compact latent representation of emissions behavior. By learning to encode only critical information and discarding unnecessary factors, the model simplifies complex nonlinear relationships into a compact latent domain where accurate predictions can be made more efficiently. To enhance computational efficiency, model compression techniques such as pruning, quantization, and low-rank factorization are applied. These strategies significantly reduce inference time and memory demand without compromising accuracy. The combined latent-space and compression framework yields a lightweight yet accurate model, providing better generalization than high-performing LSTMs while remaining suitable for real-time deployment. This work establishes a foundation for deploying advanced emission control methods, where strategies such as model predictive control and dynamic optimization can be applied effectively in hybrid and conventional powertrains.