ADVANCED MATERIALS & PROCESSES | NOVEMBER/DECEMBER 2025 18 prediction sets satisfy: P(yt+1∈ C(Xt+1)) ≥ 1−α, ensuring valid marginal coverage irrespective of distributional shifts. Calibration was improved using Platt scaling with temperature-adjusted logits, refining interval reliability across material classes. Model uncertainty was further estimated via Monte Carlo dropout (p = 0.15), where repeated stochastic forward passes provided distributional estimates of model variance. Sintering shrinkage followed an Arrhenius-type relation: where (Q) is activation energy. Semilog plots of shrinkage vs. 1000/T (as shown in Fig. 5) illustrate material- specific differences: • Al₂O₃: Q ≈ 520 kJ/mol, optimal ~1600°C. • SiC: Q ≈ 650 kJ/mol, requiring >2100°C. • 316L: Q ≈ 280 kJ/mol, densification near 1250°C. INDUSTRIAL IMPACT AND FUTURE DIRECTIONS Containerized microservices archi- tecture enables real-time deployment with single predictions executing in 12.3 ± 2.1 ms and batch optimization (50 points) in 187 ± 15 ms. Integration with Manufacturing Execution Systems through RESTful APIs facilitates incremental learning from production data. Across industrial trials, the framework reduced parameter optimization cycles by 40%, decreased build failures by 60%, and accelerated powder qualification 3×. Single-architecture deployment across 23+ material classes eliminates material-specific protocols while maintaining computational efficiency. Physics-informed neural network (PINN) development incorporates governing equations directly into loss functions through penalty terms enforcing Washburn constraints and packing bounds, improving extrapolation while maintaining physical feasibility. Autonomous experimental design via Bayesian optimization agents demonstrates 30% reduction in qualification experiments through Expected Improvement acquisition. Extension to functionally graded materials requires addressing interface physics, thermal expansion mismatch, and coupled sintering kinetics. Metal-ceramic composite studies show feasibility pending refined co-sintering models. CONCLUSIONS This physics-informed multi-agent framework transforms BJAM from empirical iteration to predictive engineering through quantile-based uncertainty quantification, physics- constrained learning, and material- adaptive processing. Validated 92% accuracy across 2847 experimental runs, 40% cycle reduction, and 60% failure decrease demonstrate industrial viability while consistently achieving ≥90% theoretical density. The modular architecture facilitates continuous adaptation to emerging materials and techniques, establishing a foundation for physics-informed manufacturing AI extending beyond BJAM to powder-based processes requiring multi-physics optimization. View an interactive demonstration available at bjampredictions.streamlit. app. ~AM&P Note: Portions of the text, editing, and figure preparation for this article were developed with the assistance of Open- AI’s ChatGPT. All technical content, interpretations, and conclusions were reviewed and approved by the author. For more information: Bhargavi Mummareddy, Knoxville, Tenn., mummareddybhargavi@gmail.com, linkedin.com/ in/bhargavi-mummareddy. Selected References 1. Y. Tao, et al., Physics-informed Machine Learning for Materials Science, Additive Manufacturing Letters, 2, 100096, 2022. 2. J. Qin, et al., Research and Application of Machine Learning for Additive Manufacturing, Additive Manufacturing, 58, 102691, 2022. Fig. 4 — Binder saturation vs. green density with binder saturation prediction intervals. Fig. 5 — Sintering shrinkage vs. temperature semi log plots showing distinct activation energies on diverse material systems, i.e., oxide, carbide, and a metallic system.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==