ADVANCED MATERIALS & PROCESSES | NOVEMBER/DECEMBER 2025 15 Binder jet additive manufacturing (BJAM) represents one of the most promising technologies for high-volume production of metal and ceramic components, offering unprecedented design freedom without the thermal stresses inherent in laser-based processes. Unlike powder bed fusion techniques, binder jetting operates at room temperature by selectively depositing liquid binder onto powder layers, enabling processing of materials difficult or impossible to handle with high-energy source methods. Its ability to simultaneously print multiple parts with complex internal geometries while maintaining dimensional accuracy positions it as a critical AM technique for aerospace, automotive, and energy applications that require both geometric complexity and material performance. Despite these advantages, BJAM faces a critical bottleneck—process parameter discovery. Current workflows rely heavily on design of experiment methodologies or a common trial and error method, requiring numerous builds to optimize combinations of particle size distribution, binder saturation, roller traverse speed, and sintering schedules. This approach scales poorly when introducing new powders or multimodal blends. The nonlinear multiphysics nature of BJAM, involving particle packing, binder capillary penetration, powder spreading dynamics, and sintering kinetics demands a fundamentally different approach. This article presents a physics- informed multi-agent AI system that operationalizes governing BJAM equations within machine learning frameworks. By decomposing optimization into specialized agents handling physics enforcement, materials classification, and uncertainty quantification, the framework captures cross-material trends while ensuring physical realism, transforming BJAM development from empirical iteration into predic- tive engineering. PHYSICS-INFORMED FOUNDATIONS The framework enforces funda- mental BJAM physics to constrain feasible design regions and prevent non-physical recommendations, including the following: Particle packing density. Random close packing of monosized spheres yields ~60% relative density, while engineered multimodal blends achieve 65-70%. The Furnas relationship follows: where ϕ represents porosity fraction. Binder saturation. Controls inter- particle adhesion and dimensional stability according to: Physics-based constraints enforce 70% ≤ S ≤ 90% to prevent delamination (S < 70%) or dimensional inaccuracy from bleeding (S > 90%). Capillary penetration. Binder penetration follows the Washburn equation: where L is penetration depth, γ is surface tension, r is effective pore radius, θ is contact angle, η is binder viscosity, and t is contact time. Spreading stability. Roller speed optimization follows empirically derived scaling: where α ≈ 0.6 for spherical powders, ensuring uniform layer formation. Sintering kinetics. Densification incorporates diffusion-controlled shrinkage: Material-specific activation energies (Q) define processing windows: Al₂O₃ (~520 kJ/mol, 1600°C), 316L (~280 kJ/mol, 1250°C), and SiC (~650 kJ/mol, >2100°C). MULTI-AGENT ARCHITECTURE The system transforms two user inputs, material type and median particle size (D₅₀), which are put into optimized BJAM process parameters through a coordinated nine-step workflow, as shown in Fig. 1. 1. Physics Agent • Computes packing density bounds (Furnas model). • Estimates binder penetration (Washburn equation). • Sets priors for layer thickness (~4 × D₅₀) and roller speed (∝ 1/D₅₀). • Applies a 0.9× correction for ceramics to account for angular morphology. 2. Constraint Manager – Pass 1 • Enforces feasibility windows: binder saturation 70-90%, sintering temperature ranges from diffusion- controlled activation energies, and printer hardware limits. • Applies binder compatibility rules: aqueous for oxides/carbides; solvent-based for metals. 3. Regression Agent • Implements quantile regression with three gradient boosting models (Q₁₀, Q₅₀, Q₉₀). • Produces conservative, median, and optimistic density predictions. • Provides actionable confidence intervals for decision-making. 4. Feature Engineering • Numerical inputs (D₅₀, thickness, saturation, roller speed) scaled. • Categorical inputs (material type, class, binder chemistry) one-hot encoded. • Interaction terms capture material- specific behavior for transfer across metals, ceramics, and carbides. 5. Materials Agent • Classifies systems and applies binder/saturation ranges. • Metals: solvent binders; saturation 70-85%; roller speed 1.2-3.5 mm/s; densities 88-95%. • Ceramics: aqueous binders; saturation 75-90%; reduced roller speed for angular particles; densities 85-92%. 6. Uncertainty Agent • Propagates errors through bootstrap sampling and cross-validation. • Performs Monte Carlo simulations (n = 10,000) to calibrate confidence intervals. • Flags predictions outside reliability thresholds.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==