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

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.

💡 Research Summary

The paper addresses the challenge of accurately predicting engine exhaust emissions during highly transient events such as rapid acceleration and deceleration, which is essential for meeting stringent environmental regulations and optimizing power‑train performance. Traditional data‑driven approaches—multilayer perceptrons (MLPs) and long short‑term memory (LSTM) networks—have shown improvements over phenomenological models but suffer from fundamental architectural limitations. They attempt to learn a direct mapping from the full high‑dimensional input space to the output space, making them sensitive to noise, dataset variability, and requiring increasingly deep networks to achieve marginal gains. Consequently, their generalization ability is limited, and their computational cost is prohibitive for on‑board deployment.

To overcome these issues, the authors propose a Joint Embedding Predictive Architecture (JEPA) that operates in a structured latent space. In JEPA, past input sequences and masked future targets are encoded separately into a common latent representation. The training objective pulls together embeddings that correspond to physically consistent past‑future pairs while pushing apart inconsistent pairs. This forces the network to discover a compact set of latent variables that capture the core dynamics of emission formation, effectively filtering out irrelevant noise and spurious correlations. The resulting latent‑space model predicts the temporal evolution of emissions (NOx, CO₂, CO, THC) from readily available ECU signals—engine torque, speed, and air‑fuel ratio—without directly modeling the full nonlinear mapping.

The dataset used for training and evaluation is uniquely comprehensive. It combines real‑world driving data collected with a Portable Emissions Measurement System (PEMS) on a BMW 530e hybrid (over 10 hours of operation, 146 channels sampled at 5 Hz) with high‑frequency hardware‑in‑the‑loop (HIL) bench measurements of the same B48 engine. The on‑road profiles are reproduced on a test bench, providing both realistic transient excursions and high‑resolution steady‑state points (66 evenly spaced operating points). This dual‑source data ensures that the JEPA learns both the stochastic nature of real driving and the deterministic physics captured in controlled experiments.

Experimental results compare the JEPA against the best‑performing LSTM baseline from prior work, using identical training‑test splits and evaluation metrics. Across all target species, JEPA achieves lower mean absolute error, especially in capturing rapid peaks, phase shifts, and amplitude variations during transients. Data efficiency is also demonstrated: JEPA reaches comparable accuracy with fewer training samples, indicating a more robust representation.

Recognizing the need for real‑time on‑board inference, the authors apply model compression techniques. Structured pruning removes entire channels and layers based on sensitivity analysis, preserving the dimensional consistency required for the encoder‑decoder pipeline. Post‑training quantization to 8‑bit integer precision further reduces memory footprint and bandwidth without significant loss of fidelity. After compression, the model’s inference latency drops by more than 30 % and its memory usage is cut by a factor of four, while the average error increase remains below 2 % for all emissions.

The paper’s contributions can be summarized as follows: (1) introduction of a latent‑space JEPA framework that learns disentangled, physics‑aware embeddings for transient emission dynamics; (2) creation of a rich, dual‑source dataset that bridges real‑world variability and high‑resolution bench data; (3) systematic evaluation showing superior accuracy and generalization over state‑of‑the‑art LSTM models; and (4) demonstration of practical deployment through structured pruning and quantization, yielding a lightweight model suitable for integration into ECUs for model‑predictive control or model‑based reinforcement learning.

Overall, the work provides a clear pathway toward more robust, data‑efficient emission modeling that can be directly embedded in next‑generation vehicles, enabling proactive control strategies that reduce regulated pollutants while preserving fuel efficiency.