A Decision Transformer trained on human gameplay optimizes grain boundary networks, reaching about 92% of simulated annealing's solution quality with orders of magnitude fewer iterations and transferring across material models without retraining.
Atomic-scale grain boundary engineering to overcome hot-cracking in additively-manufactured superalloys
1 Pith paper cite this work. Polarity classification is still indexing.
abstract
There are still debates regarding the mechanisms that lead to hot cracking in parts build by additive manufacturing (AM) of non-weldable Ni-based superalloys. This lack of in-depth understanding of the root causes of hot cracking is an impediment to designing engineering parts for safety-critical applications. Here, we deploy a near-atomic-scale approach to investigate the details of the compositional decoration of grain boundaries in the coarse-grained, columnar microstructure in parts built from a non-weldable Ni-based superalloy by selective electron-beam melting. The progressive enrichment in Cr, Mo and B at grain boundaries over the course of the AM-typical successive solidification and remelting events, accompanied by solid-state diffusion, causes grain boundary segregation induced liquation. This observation is consistent with thermodynamic calculations. We demonstrate that by adjusting build parameters to obtain a fine-grained equiaxed or a columnar microstructure with grain width smaller than 100 $\mu$m enables to avoid cracking, despite strong grain boundary segregation. We find that the spread of critical solutes to a higher total interfacial area, combined with lower thermal stresses, helps to suppress interfacial liquation.
citation-role summary
citation-polarity summary
fields
cond-mat.mtrl-sci 1years
2024 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
A Decision Transformer Approach to Grain Boundary Network Optimization
A Decision Transformer trained on human gameplay optimizes grain boundary networks, reaching about 92% of simulated annealing's solution quality with orders of magnitude fewer iterations and transferring across material models without retraining.