Close-in exoplanets span a wide range of orbital architectures and physical properties, consistent with the combined influence of migration processes and formation conditions. Although population synthesis models predict the emergence of distinct planetary populations, establishing a statistically robust connection between observed samples and synthetic populations remains challenging, particularly in high-dimensional parameter spaces dominated by dynamical quantities. Intrinsic population-level organisation within the observed sample of close-in exoplanets is investigated through physically motivated parameters, with the aim of establishing a quantitative link to formation pathways predicted by pebble-accretion models. A two-stage Gaussian mixture model (GMM) was applied to an observed sample of close-in exoplanets, with unsupervised probabilistic clustering performed in a feature space dominated by dynamical descriptors of planet--star interactions. The resulting clusters were mapped onto a pebble-accreted synthetic population within a statistically motivated 3D parameter space. Formation-related quantities, including formation timing, gas fraction, and ice--rock mass ratio, were then employed to establish a probabilistic interpretation and comparison of the mapped clusters. Statistically supported sub-populations were identified without the imposition of predefined classification boundaries, including very-massive gas giants, hot giants, warm-Jupiter-dominated systems, and lower-mass giants. When mapped onto pebble-accreted synthetic populations, these observational clusters exhibited systematic differences in formation timing, gas fraction, and solid growth histories across the dominant populations. Data-driven clustering in a dynamical-parameter space enables a direct and statistically robust comparison between observed exoplanet sub-populations and pebble-accretion synthetic populations. The inferred cluster-level trends point to systematic differences in formation timing and accretion histories among giant populations, with very-massive gas giants preferentially associated with earlier formation epochs than hot-giant and warm-Jupiter-dominated systems. This demonstrates the potential of physically motivated machine-learning approaches to connect observed exoplanets with theoretical formation scenarios at the population level.