NeuralMOVES: A lightweight and microscopic vehicle emission estimation model based on reverse engineering and surrogate learning

AbstractThe transportation sector accounts for nearly one-quarter of global greenhouse gas (GHG) emissions. Emerging technologies-such as eco-driving, connected vehicle control, and others-offer significant potential for emission reduction; however, officially validated, yet optimization-ready emission models are essential for guiding their design, deployment, and evaluation. The U.S. EPA’ Motor Vehicle Emission Simulator (MOVES) is the validated regulatory and industry standard for vehicle emissions in the U.S. Yet, its complexity, macroscopic focus, and high computational demands make it unsuitable and incompatible with control and optimization applications, and burdensome even for traditional analyses. Furthermore, its reliance on location-specific inputs limits its applicability beyond the U.S. As a result, researchers often resort to alternative models, leading to emission estimates that are neither comparable nor officially validated. To address this gap, we introduce NeuralMOVES, an open-source, lightweight surrogate model for CO₂ emissions with near-MOVES fidelity. NeuralMOVES transforms MOVES from a multi-software system requiring specialized expertise and hours of computation into a 2.4 MB Python package that runs in milliseconds and integrates seamlessly into optimization frameworks. Developed by reverse-engineering MOVES through over 200 million batch queries to generate a comprehensive microscopic emission dataset (MOVES_RE, 9.89 GB), NeuralMOVES uses machine learning to compress this dataset by over 4,000× while enabling continuous, differentiable, and real-time emission estimation. An extensive validation shows a mean absolute percentage error of 6.013% across over two million test scenarios, each representing a complete driving trajectory evaluated under specific environmental and vehicle conditions. We demonstrate NeuralMOVES in a dynamic eco-driving case study, showing that it integrates seamlessly into optimization pipelines, leads to different trajectories than alternative models, and captures parameter sensitivities that alternative models overlook. NeuralMOVES enables regulatory-grade, microscopic emission modeling for emerging transportation technologies worldwide and is available at: https://github.com/edgar-rs/neuralMOVES.