State-enhanced attention network for optimisation of energy and yield in gas atomised metal powder production

Abstract

Gas atomisation is a widely used technique for producing spherical metal powder feedstock for additive manufacturing. However, the process parameters suffer from variability and inefficiency in balancing powder yield, energy consumption, and particle size distribution. Optimising these complex, interdependent parameters pose a significant challenge. This work proposes a novel State-Enhanced Attention Network architecture in a framework that simultaneously optimises yield and energy consumption during nitrogen gas atomisation for sustainable metal powder production. The novelty lies in integrating processed long-term memory states with the attention mechanism, enabling nuanced attention weighting. This allows the model to leverage global sequence context and recent state information for improved yield and energy predictions. The proposed network is trained and integrated into a non-dominated sorting genetic algorithm to enable multi-objective optimisation. This framework evolves a set of Pareto-optimal solutions that balance trade-offs between maximising yield and minimising energy consumption. The approach is evaluated using augmented real-world data from an industrial gas atomisation plant. The proposed model demonstrates significantly improved predictive accuracy on real-world datasets, compared with baseline deep learning models. Results highlight the capabilities of the proposed technique for automated, data-driven optimisation of gas atomisation, simultaneously improving yield, energy efficiency, quality control, and sustainability. The integrated deep learning and evolutionary optimisation framework also provides an innovative solution for enhanced control of additive manufacturing powder production processes.