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

TL;DR

The paper tackles the challenge of accurately predicting transient engine emissions under rapid operating changes, where conventional monolithic neural nets struggle to generalize. It introduces a latent-space framework based on Joint Embedding Predictive Architecture (JEPA) that encodes past emissions and future inputs into a compact latent representation, enabling a predictive latent dynamics model trained with VICReg and cross-covariance penalties. Compared with a baseline LSTM, JEPA delivers superior data efficiency and fidelity during transients, while enabling real-world deployment through structured pruning and post-training quantization, with decoders providing domain-recovery interpretability. The work demonstrates the potential of latent-space, model-compression-friendly emission models for on-board predictive control in conventional and hybrid powertrains, and outlines future directions including richer sensing, physics-informed regularization, and hardware-aware compression.

Abstract

Accurately modeling and controlling vehicle exhaust emissions during transient events, such as rapid acceleration, is critical for meeting environmental regulations and optimizing powertrains. Conventional data-driven methods, such as Multilayer Perceptrons (MLPs) and Long Short-Term Memory (LSTM) networks, improve upon phenomenological models but often struggle with the complex nonlinear dynamics of emission formation. These monolithic architectures are sensitive to dataset variability and typically require deep, computationally expensive structures to perform well, limiting their practical utility. This paper introduces a novel approach that overcomes these limitations by modeling emission dynamics within a structured latent space. Leveraging a Joint Embedding Predictive Architecture (JEPA), the proposed framework learns from a rich dataset that combines real-world Portable Emission Measurement System (PEMS) data with high-frequency hardware-in-the-loop measurements. The model abstracts away irrelevant noise, encoding only the key factors governing emission behavior into a compact, robust representation. This results in superior data efficiency and predictive accuracy across diverse transient regimes, significantly outperforming high-performing LSTM baselines in generalization. To ensure suitability for real-world deployment, the JEPA framework is structured to support pruning and post-training quantization. This strategy drastically reduces the computational footprint, minimizing inference time and memory demand with negligible accuracy loss. The result is a highly efficient model ideal for on-board implementation of advanced strategies, such as model predictive control or model-based reinforcement learning, in conventional and hybrid powertrains. These findings offer a clear pathway toward more robust emission control systems for next-generation vehicles.

Figures (9)

• Figure 1: On-road speed–torque profile for dataset generation. The envelope and transient excursions derived from this profile are used to define bench trajectories and to focus measurements in emission-critical regions.
• Figure 2: Test bench for dataset generation. Raw exhaust is sampled between the turbocharger and the catalytic converter (red arrow), allowing for transient-resolved emissions measurements upstream of aftertreatment systems.
• Figure 3: Schematic of the test bench setup. The engine is coupled to an electric machine (load) via the torque flange and a central measurement system. Standard thermocouple and pressure sensors are omitted here for clarity.
• Figure 4: Schematic representation of the JEPA training and inference process. Illustrates how future input sequences and past emission observations are jointly encoded into a latent space. The model predicts future latent states via autoregressive dynamics, using multiple loss functions (variance, invariance, covariance, and cross-covariance) to regularize latent representations. Encoders, predictor network, and losses are highlighted for clarity. Arrows denote the sequence of information flow through encoders $h_\eta$, $g_\phi$, and the predictor $f_\theta$ for emission forecasting across time horizons.
• Figure 5: Schematic representation of state and emission encoder–decoder. Diagram illustrating the relationships between engine operating variables ($\tau$, $\omega$, $\lambda$) and emission outputs (NO$_x$, CO$_2$, CO, THC), via their latent embeddings. Separate encoders ($g_\phi$ and $h_\eta$) transform input and emission sequences into their respective latent representations. The corresponding decoders ($g^{-1}_\phi$ and $h^{-1}_\eta$) reconstruct the original sequences from the latent space. This architecture enables reconstruction-based validation and interpretability of the learned latent dynamics, facilitating performance evaluation and closed-loop control design.