Physics-Informed Neural Networks for Vehicle Emissions: How the Technology Works

Conventional approaches to vehicle emissions calculation fall broadly into two categories: empirical models that apply statistical averages to fleet data, and hardware sensors that measure exhaust gases directly. Physics-Informed Neural Networks (PINNs) represent a third approach — one that embeds the governing physical equations of vehicle dynamics directly into a machine learning architecture.

The Core Concept

A standard neural network learns patterns from data. Given enough examples of inputs and outputs, it can approximate the relationship between them. The limitation for emissions calculation is that learning from data alone — without physical constraints — can produce outputs that are statistically plausible but physically impossible.

A Physics-Informed Neural Network adds a second objective to the training process. In addition to fitting the observed data, the network must simultaneously satisfy the physical equations that govern the system being modelled. For a heavy goods vehicle, these equations describe how the forces acting on a vehicle — traction, aerodynamic drag, gradient resistance, rolling resistance — determine its fuel consumption at every point along a journey.

The result is a model that respects the physics of vehicle dynamics. It cannot, for example, infer a fuel consumption value that violates conservation of energy. The physical equations act as hard constraints that bound the solution space to what is physically possible.

Why This Matters for CO2e Calculation

For HGV emissions calculation, the PINN approach offers a specific advantage: the ability to calibrate vehicle-specific physical parameters — engine efficiency (η) and rolling resistance coefficient (C_rr) — directly from operational telemetry, without requiring manual measurement or manufacturer data.

These parameters vary between vehicles of nominally the same type. An older engine operating under heavy load in hilly terrain will have different effective efficiency characteristics than a new engine on a flat motorway run. Standard emission factor approaches cannot capture this variation. A PINN-based system, by solving for these parameters as part of the training process, produces calculations that are specific to each vehicle's actual physical characteristics.

The State of the Research

PINNs were formalised as a computational methodology by Raissi, Perdikaris, and Karniadakis (2019) and have since been applied across a range of scientific and engineering domains — fluid dynamics, heat transfer, structural mechanics, and others. Application to vehicle dynamics and transport emissions calculation is an active area of research.

EcoTrace is developing a PINN-based architecture specifically for HGV fuel consumption and CO2e calculation — the Scientific Carbon Validation Engine (SCVE). Proof-of-concept experiments conducted in early 2026 on synthetic datasets achieved a Mean Absolute Percentage Error (MAPE) of 0.62% for fuel consumption prediction on a variable-gradient route, within the sub-1% target specified for the system.

This work is at the R&D stage. The full system — incorporating real vehicle telemetry, multi-fleet calibration, and ISO 14083 alignment — is the subject of ongoing development.

Practical Implications

For logistics operators and sustainability teams, the practical implication of a PINN-based approach is the possibility of CO2e calculation accuracy approaching hardware measurement, delivered as software that processes existing telemetry data — without requiring physical sensors to be installed on every vehicle.

Whether this promise is fully realised in operational conditions with heterogeneous fleets and real-world data quality is the central technical question that EcoTrace's R&D programme is designed to answer.