Empirical Kinetic Modeling Approach (EKMA)

• EKMA is a data-driven framework that models system evolution through kinetic interactions among key variables, such as ozone precursors and traffic densities.
• It leverages empirical and machine learning surrogate models to calibrate kinetic coefficients, achieving high predictive accuracy with metrics like R² and RMSE.
• The approach facilitates regime classification and sensitivity analysis, informing practical strategies in air quality management and traffic engineering.

The Empirical Kinetic Modeling Approach (EKMA) is a data-driven framework for diagnosing and modeling the nonlinear dynamics of complex systems where mechanistic details are either unknown or computationally intractable. Originating in atmospheric chemistry for interpreting ozone formation regimes, EKMA conceptualizes system evolution as emergent from interactions between key precursors or compartments, parameterized via empirical or fitted kinetic laws. Recent developments extend EKMA beyond mechanistic chemistry, leveraging statistical and machine learning surrogates for modeling atmospheric processes, and cast kinetic flow models in traffic engineering as compartmental reaction networks, enabling direct parameter calibration from high-dimensional observational data (Zheng, 18 Jan 2026, Pereira et al., 2021).

1. Conceptual Foundations of EKMA

EKMA models the evolution of a system as a set of kinetic interactions among primary reactants or state variables, representing these in either analytical or algorithmic form. In atmospheric chemistry, the canonical EKMA framework visualizes photochemical ozone production as a function of its main precursors—typically nitrogen oxides (NOₓ) and volatile organic compounds (VOCs)—under varying environmental conditions. These dependencies are represented as response surfaces or isopleths in the precursor space, delineating regions of distinct regime behavior, such as NOₓ-limited or VOC-limited ozone formation (Zheng, 18 Jan 2026).

In traffic modeling, EKMA formalizes vehicle movement as a discrete compartmental process, with road segments as nodes (compartments) and fluxes between them treated as chemical-like reactions, as in the discrete kinetic formulation for the LWR (Lighthill–Whitham–Richards) traffic flow model (Pereira et al., 2021). Here, kinetic coefficients (e.g., reaction rates or transfer rates) connect compartmental densities to propagating fluxes, which are fit empirically to reproduce observed data.

2. Machine Learning–Based Surrogate EKMA Frameworks

A recent EKMA evolution replaces explicit chemical or physical laws with data-driven surrogate models. In Zheng (2024–2025), urban ozone concentrations are estimated using a random forest (RF) regression trained on hourly measurements of precursors and spatiotemporal features. The model input vector at time $t$ is

$$X_t = \left(\mathrm{NO}_2{}_t, \mathrm{CO}_t, \mathrm{PM}_{2.5_t}, \mathrm{lon}_t, \mathrm{lat}_t, \sin\left(\frac{2\pi\, \mathrm{hour}_t}{24}\right), \cos\left(\frac{2\pi\, \mathrm{hour}_t}{24}\right), \ldots\right)^T,$$

producing surrogate estimates

$\widehat{[\mathrm{O}_3]}_t = f_{\mathrm{RF}}(X_t).$

In the cited Los Angeles study, the surrogate achieved $R^2 = 0.857$ and RMSE = 0.006 ppb on the test set, capturing 86% of out-of-sample ozone variance. The model's predictive structure allows perturbation analyses analogous to classical EKMA sensitivity experiments, but with the advantage of utilizing continuous and high-dimensional observational data rather than relying on mechanistic or heavily parameterized chemistry models (Zheng, 18 Jan 2026).

3. Quantifying Influence and Sensitivities

Permutation importance is applied post hoc to the trained surrogate to quantify the relative influence of input features on predicted outcomes. In the Los Angeles framework, the dominant predictors were diurnal cycle features (12.3%, 11.8%), followed by NO₂ (10.5%), CO (7.8%), and PM₂.₅ (5.2%). Spatiotemporal encodings for month and weekday provided additional explanatory power (≈4–5%) (Zheng, 18 Jan 2026). This ordering confirms the interplay between precursor availability and temporal patterns in modulating ozone dynamics.

EKMA-style sensitivity is diagnosed by perturbing precursor concentrations:

$$\text{NO}_2{}_i^{(\alpha)} = \alpha \mathrm{NO}_2{}_i, \quad \mathrm{CO}_i^{(\beta)} = \beta \mathrm{CO}_i,$$

for scaling factors $\alpha, \beta \in [0.5, 1.5]$. The response surface

$$\bar O_3(\alpha, \beta) = \frac{1}{|\mathcal{B}|} \sum_{i \in \mathcal{B}} f_{\mathrm{RF}}(X_i^{(\alpha, \beta)}),$$

is evaluated on a baseline sample set $\mathcal{B}$ (summer afternoons with high ozone), and finite differences compute sensitivities:

$$S_{\mathrm{NO}_x}(\alpha, \beta) \approx \frac{\bar O_3(\alpha+\Delta\alpha,\beta) - \bar O_3(\alpha,\beta)}{\Delta\alpha},$$

$$S_{\mathrm{VOC}}(\alpha, \beta) \approx \frac{\bar O_3(\alpha,\beta+\Delta\beta) - \bar O_3(\alpha,\beta)}{\Delta\beta},$$

with $\Delta\alpha = \Delta\beta = 0.01$. Regions where $S_{\mathrm{VOC}} > S_{\mathrm{NO}_x}$ are VOC-limited and vice versa (Zheng, 18 Jan 2026).

4. EKMA Diagrams and Regime Classification

Contour diagrams (isopleths) are constructed by evaluating $\bar O_3(\alpha, \beta)$ on a dense $\alpha$–$\beta$ grid and plotting level sets for constant ozone concentration. The regime-transition curve, marking $S_{VOC} = S_{NO_x}$, is obtained via bilinear interpolation; points above this curve are VOC-limited, while those below are NOₓ-limited.

For the Los Angeles case, the response surface demonstrates that VOC reductions (proxied by CO) are more effective than NO₂ reductions across most of the relevant precursor space: for a 10% decrease in CO alone, $\Delta \mathrm{O}_3 \approx -1.5 \,\mathrm{ppb}$, versus $-0.8\,\mathrm{ppb}$ for an equivalent NO₂ cut. The transition curve generally follows $\beta \approx \alpha$ near unity, implying that net ozone decreases under typical summer-afternoon conditions require a policy emphasis on VOC controls (Zheng, 18 Jan 2026).

5. Kinetic Compartment Models in Traffic

The EKMA methodology has been extended to traffic flow, where roadways are discretized into $N_x$ cells and vehicle density is viewed as a reactant concentration. Vehicle transfers are modeled as chemical reactions:

$$\Phi_j + O_{j-1} \xrightarrow{k_{j-1\to j}} O_j + \Phi_{j-1},$$

with $\Phi_j$ as “free” space ($\rho_m - \rho_j$) and $O_j$ as “occupied” space ($\rho_j$). Derivation via the law of mass action yields compartmental ODEs for densities:

$$\frac{d\rho_j}{dt} = k_{j-1\to j} \rho_{j-1}(\rho_m-\rho_j) - k_{j\to j+1}\rho_j(\rho_m-\rho_{j+1}).$$

Normalizing and discretizing gives the Traffic Reaction Model (TRM):

$$U_j^{n+1} = U_j^n + C_j^n U_{j-1}^n(1-U_j^n) - C_{j+1}^n U_j^n(1-U_{j+1}^n),$$

where $C_j^n$ are normalized reaction coefficients (related to maximal speed $v_m$), and $U_j^n$ normalized densities (Pereira et al., 2021). The parameter-estimation problem focuses on calibrating these coefficients to match observed densities, using nonlinear least-squares and regularization to mitigate overfitting. The forward–backward “back-propagation” strategy enables efficient gradient computation.

6. Applications, Extensions, and Limitations

Applications of EKMA span atmospheric regime diagnosis, real-time traffic calibration, and surrogate modeling in domains with partial observations or high-dimensional parameter spaces. In ozone control, EKMA-style analysis provides empirical support for prioritizing VOC emission reductions in urban air quality management when the regime is VOC-limited (Zheng, 18 Jan 2026). In traffic, fitting reaction rates to observed density profiles enables detailed reconstruction of capacity drops and congestion transitions, with error metrics such as RMSE and relative $v_m$ error reporting successful validation on synthetic and real-world highway data (Pereira et al., 2021).

Extensions include reaction-graph generalization for urban traffic networks, robust performance under partial compartment observations, and kinetic models with data-driven rate functions or machine-learning surrogates. However, limitations include potential overfitting due to high-dimensional parameterizations, CFL constraints on time–space resolution, and model misspecification for multimodal or highly nonlinear dynamics.

A plausible implication is that EKMA frameworks, particularly those leveraging machine learning surrogates, will continue to expand as computational resources and data availability increase, offering interpretable and updatable alternatives to conventional mechanistic models in environmental diagnostics and infrastructure analysis (Zheng, 18 Jan 2026, Pereira et al., 2021).