Developing sustainable pathways for metal processing is increasingly critical. While recycling can contribute to sustainability, it also introduces variability in material composition, requiring on-demand optimization of processing conditions. Traditional modelling of processing conditions and microstructure linkages often produce high fidelity outputs that face computational bottlenecks with limited flexibility and explainability to capture statistical relationship between variables. To address this, we present a conditional generative machine learning (ML) framework leveraging GPU-accelerated normalizing flow models for process-structure linkages in steels. Conditioned on tabular processing parameters---cooling schedules and alloying compositions---the model infers the posterior distributions of microstructure statistics---including global quantities such as phase fraction and interface area fraction and grain-level metrics such grain size and grain aspect ratio. Trained on a compact dataset, the model accurately captures statistical interdependencies, showing strong agreement between predicted and true posterior distributions. We further employ SHapley Additive exPlanations (SHAP) analysis for explainability and measure the contribution of individual processing parameters towards global and grain-specific microstructure statistics. This approach offers a promising avenue for intelligent, adaptive material design and process optimization across diverse material classes.