Machine learning assisted design of high entropy alloys with desired property

We formulate a materials design strategy combining a machine learning (ML) surrogate model with experimental design algorithms to search for high entropy alloys (HEAs) with large hardness in a model Al-Co-Cr-Cu-Fe-Ni system. We fabricated several alloys with hardness 10% higher than the best value i...

Celý popis

Uložené v:

Podrobná bibliografia
Vydané v:	Acta materialia Ročník 170; číslo C; s. 109 - 117
Hlavní autori:	Wen, Cheng, Zhang, Yan, Wang, Changxin, Xue, Dezhen, Bai, Yang, Antonov, Stoichko, Dai, Lanhong, Lookman, Turab, Su, Yanjing
Médium:	Journal Article
Jazyk:	English
Vydavateľské údaje:	United States Elsevier Ltd 15.05.2019 Elsevier
Predmet:	Active learning Machine learning Materials genome initiative Multi-principal element alloys Materials genome initiative Multi-principal element alloys Active learning Machine learning
ISSN:	1359-6454, 1873-2453
On-line prístup:	Získať plný text
Tagy:	Pridať tag Žiadne tagy, Buďte prvý, kto otaguje tento záznam!

Popis
Shrnutí:	We formulate a materials design strategy combining a machine learning (ML) surrogate model with experimental design algorithms to search for high entropy alloys (HEAs) with large hardness in a model Al-Co-Cr-Cu-Fe-Ni system. We fabricated several alloys with hardness 10% higher than the best value in the original training dataset via only seven experiments. We find that a strategy using both the compositions and descriptors based on a knowledge of the properties of HEAs, outperforms that merely based on the compositions alone. This strategy offers a recipe to rapidly optimize multi-component systems, such as bulk metallic glasses and superalloys, towards desired properties. In the present study, we proposed a data-driven approach combining machine learning, experimental design and feedback from experiment to accelerate the search for multi-component alloys with target properties. We demonstrated the efficiency of our approach by rapidly obtaining several alloys with hardness 10% higher than the best value in the original training dataset via only seven experiments. In Iteration Loop I, a machine learning surrogate model is trained to learn the property-composition, relationship, yi=f(ci), with associated uncertainties. The model is applied to the vast unexplored space and a utility function is employed to recommend the most informative candidate for the next experiment, which balances the exploitation and exploration. Feedback from experimental synthesis and characterization allows the subsequent iterative improvement of the surrogate model. Iteration Loop II is essentially same as Iteration Loop I, except that a features pool was introduced to the Iteration Loop I and a surrogate model is trained from composition (ci) and the preselected physical features (pi), yi=f(ci,pi). We found that the approach using both the composition and the descriptors based on domain knowledge can more effectively accelerate material optimization compared to the approach using only the compositions. [Display omitted]
Bibliografia:	USDOE
ISSN:	1359-6454 1873-2453
DOI:	10.1016/j.actamat.2019.03.010