Modeling complex powder and particle processes often requires a compromise between the interpretability of physics-based mechanistic models and the predictive performance of data-driven methods. While “white-box” models provide physical insight, they are frequently limited by complexity and computational cost. Conversely, “black-box” approaches (e.g., deep learning) can achieve high accuracy but may generalize poorly and violate fundamental physical constraints. This work presents a gray-box hybrid modeling workflow augmented by an AI Research Agent to accelerate model development and reduce expert effort in configuring advanced algorithms. The modular framework integrates physics-inspired structure (e.g., Dimensional Analysis) with Symbolic Regression / Genetic Programming for compact, interpretable equation discovery, and optionally Graph Neural Networks to capture relational structure in particle/process data. To overcome the “expert bottleneck” in method selection, hyperparameter setup, and constraint formulation, we introduce an AI Research Agent built on Large Language Models (LLM) with Retrieval-Augmented Generation (RAG). The agent ingests domain knowledge from publications and technical documentation, converting unstructured content into a structured knowledge base that guides the modeling workflow. Using a two-stage configuration strategy, the agent (i) proposes and refines search spaces and parameters for equation discovery and (ii) performs indirect configuration by suggesting physics-informed penalties and constraints (e.g., dimensional homogeneity, conservation-consistent forms, boundedness, or monotonic trends) derived from retrieved knowledge. This promotes models that are not only statistically accurate but also physically consistent and interpretable. Across case studies in particle technology spanning simulation- and experiment-driven settings, the proposed approach identifies parsimonious, physically grounded expressions that match or exceed empirical baselines while improving robustness under limited data. The framework supports transferable modeling of particle-based processes and aligns with AI-driven computational approaches for manufacturing-relevant applications.