海歸學者發起的公益學術平臺
分享資訊,整合資源
交流學術,偶爾風月


圖 1元素特徵遷移對抗網路框架結構
來自中國清華大學物理系的倪軍教授團隊和日本東北大學Ying Chen教授團隊,提出了元素特徵遷移對抗網路框架,可使用生成對抗網路從原始資料集中生成可用於資料增強的新的元素特徵和材料性質,從而提高圖神經網路在小樣本資料集下的效能。元素特徵遷移對抗網路使用元素卷積圖神經網路從晶體的原子和結構資訊中提取適合描述預測目標的元素特徵。這些特徵將被轉移到生成對抗網路中,該模型用於生成具有預測目標的新元素特徵。同時,生成的元素特徵可用作輸入在廣闊的成分空間中探索材料性質,避免了結構計算帶來的龐大計算量。該研究中引入了迭代方法以提高資訊最大化生成對抗網路的精度:使用多層感知器對資訊最大化生成對抗網路生成的特徵進行性質預測,將這些結果重新輸入資訊最大化生成對抗網路中訓練,以提高網路對高熵合金性質的預測精度。

圖 2元素特徵遷移對抗網路對CrFeNiCoMn/Pd合金總能量、形成能、混合能、磁矩和方均位移的預測效能
將元素特徵遷移對抗網路應用於高熵合金的性質預測中,使用生成網路的資料增強顯著提高了模型的預測精度。模型預測了FeCoNiCrMn/Pd合金在面心立方和體心立方結構下的形成能、磁矩和自由能。該合金更傾向於形成面心立方,在普遍有著較高的磁矩的情況下,兼顧了合金強度。透過該生成模型,探索了廣闊成分空間下高熵合金的穩定性,豐富了高熵合金相關的物性研究。同時,該研究開發的元素特徵遷移對抗網路框架可用於各種多組元材料以解決機器學習資料量不足的困難,輔助材料開發與獲得新的物理知識。

圖 3系統的形成能、磁矩和自由能預測。
該文近期發表於npj Computational Materials11, 54 (2025),英文標題與摘要如下,點選左下角“閱讀原文”可以自由獲取論文PDF。
EFTGAN: Elemental features and transferring corrected data augmentation for the study of high-entropy alloys
Yibo Sun, Cong Hou, Nguyen-Dung Tran, Yuhang Lu, Zimo Li, Ying Chen & Jun Ni
Using machine learning to predict and design materials is an important mean of accelerating material development. One way to improve the accuracy of machine learning predictions is to introduce material structures as descriptors. However, the complexity of computing material structures limits the practical use of these models. To address this challenge and improve prediction accuracy in small data sets, we develop a generative network framework: Elemental Features enhanced and Transferring corrected data augmentation in Generative Adversarial Networks (EFTGAN). Combining the elemental convolution technique with Generative Adversarial Networks (GAN), EFTGAN provides a robust and efficient approach for generating data containing elemental and structural information that can be used not only for data augmentation to improve model accuracy, but also for prediction when the structures are unknown. Applying this framework to the FeNiCoCrMn/Pd high-entropy alloys, we successfully improve the prediction accuracy in a small data set and predict the concentration-dependent formation energies, lattices, and magnetic moments in quinary systems. This study provides a new algorithm to improve the performance and usability of deep learning with structures as inputs, which is effective and accurate for the prediction and development of materials for small data sets.

擴充套件閱讀
媒體轉載聯絡授權請看下方
