UNiCO-AM: Uncertainty-aware Neural Co-design for Additive Manufacturing via an Attention-based Spatio-Temporal Neural Operator and Bayesian Optimization

Metal Additive Manufacturing (AM) is a complex process where the interaction between material and process parameters can significantly influence melt-pool dynamics, thermal gradients, and solidification behavior, ultimately determining the resulting microstructure, mechanical properties, and overall quality of the fabricated part. However, inherent challenges such as the high-dimensional process-material design space, strong nonlinearity in thermal-physical interactions, and the high computational cost of high-fidelity simulations make it difficult to systematically explore and optimize these interdependencies across diverse alloys and process conditions. To address this challenge, we propose UNiCO-AM (Uncertainty-aware Neural Co-design for Additive Manufacturing), a physics-informed optimization framework that integrates an attention-based spatio-temporal neural operator (ASNO) with Bayesian optimization (BO). In this framework, we develop a material–process-agnostic ASNO model that is capable of learning the coupled spatio-temporal thermodynamics of the additive manufacturing process, accurately predicting full-field temperature evolution across different alloys and laser parameters, and generalizing to unseen material–process combinations. Another key contribution of this work is the quantification of epistemic uncertainty in ASNO using a Linear Laplace-based Approximation (LLA) method. This approach estimates the model’s uncertainty around its learned parameters, providing the confidence information required for effective BO. Incorporating uncertainty quantification not only enables more reliable and data-efficient optimization but also enhances the generalization capability of ASNO across diverse materials and process conditions. Finally, we complete the UNiCO-AM framework by establishing a concurrent material-process BO loop for AM, incorporating a q -Expected Hypervolume Improvement (q-EHVI) strategy to efficiently explore the high-dimensional design space under physical and thermal constraints. The primary objective of the UNiCO-AM framework is to enhance the strength of the fabricated part by promoting finer microstructural grain sizes through joint optimization of process parameters and material properties that regulate cooling rates and solidification behavior during AM. This approach yields a Pareto front of material–process combinations that balance cooling efficiency and thermal stability, leading to improved mechanical performance in fabricated parts while also enabling the discovery of novel material-process property combinations.