以前饋式神經網路訓練鐵鈷鎳鉻高熵合金勢能應用於微結構原子尺度模擬

Abstract

高熵合金是一種近幾年才開始快速發展的特殊合金，有別於傳統合金皆是由一種主要金屬元素，參雜其他元素所製作而成，高熵合金含有四種以上的金屬元素，且每種元素比例不超過35%。高熵合金最廣為人知的就是其四大特性，第一為高熵效應，使得各種元素隨機排列但產生結構簡單的固溶相，而不是質脆的生成物；第二為晶格扭曲效應，此效應導致高熵合金有特別的熱電性質，例如導熱速度慢，導電性差；第三為擴散延遲效應，這個使得高熵合金有效好的熱穩定性(耐高溫氧化)；最後最有潛力的雞尾酒效應，可以添加其他性質的合金使高熵合金有添加物的性質，例如添加密度低的合金能使高熵合金重量變輕。本研究嘗試以分子動力學的方式計算鐵鈷鎳鉻高熵合金的材料性質。 對於分子動力學來說，由勢能場得到結構系統的能量與受力，進而取得其運動軌跡，因此，勢能場的好壞決定了模擬的準確度。第一原理計算能獲得最為準確的作用力與勢能場。但是第一原理的計算必須耗費相當大的計算資源以及時間，並且計算尺度被限縮(幾百顆原子)，在考慮同時出現多種缺陷結構或隨機排列時，計算規模往往超過第一原理能夠負荷的量。因此，若是能找到一個方法能建立精準快速，且可以處理大尺度的勢能，對於分子模擬這領域將會有所貢獻。本研究採用機器學習的方式擬合高熵合金的勢能，隨機的原子排列方式使得訓練十分具有挑戰性，並討論將原子間的力列入訓練目標的重要性。 本研究將藉由計算高熵合金的基本性質，驗證機器學習方法訓練出的勢能與第一原理計算結果十分吻合(約1-5%的誤差)，但會受到訓練集內結構種類所限制，接著利用此勢能進行了兩千多顆原子的蒙地卡羅方法，找出各種可能存在的局部最低點，並測試此結構在室溫300K的能量與溫度穩定性，模擬結果顯示兩者皆穩定，並將比較兩者的速度，證明其具有高度精準且計算速度大幅高越第一原理。

High Entropy Alloys (HEA) is a novel metallic material that has drawn increasing attentions from both academia and industries in recent years. In contrast to conventional alloys comprising of at most one or two principal elements, HEA system consists of more than four constituent elements with concentrations ranging between 5% and 35%. The entropic mixing effect arising from high configuration entropy ensures the HEAs to remain the simple solid solution phase with FCC, BCC or HCP crystal structures without phase segregations. The lattice distortion effect from atomic size differences makes HEA has unique thermal, electric, and mechanical properties such as low thermal/electric conductivities. Lattice distortion also slows down defect diffusion, brings obstacles to dislocation glides and prevents oxidation at elevated temperatures. Finally, the cocktail effect, namely, mixing atoms of distinct properties together, allows tuning material properties of HEAs with almost infinite degrees of freedom. These aforementioned features make HEAs emerge as the rising star as the future structural and electric materials. In this thesis, we attempted to investigate the material properties of Fe-Co-Ni-Cr HEA by using machine-learning-enabled atomistic simulations. In atomistic simulations, the first principle calculations provide the most accurate system energies and atomic forces. However, the first principle calculations are computationally expensive, thereby preventing performing exhaustive sampling of configurations of HEAs. Classical molecular dynamics simulations (MD) allow efficient exploration of atomistic configurational space; however, the accuracy of MD calculations critically relies on the empirical force fields (or, potentials). In this thesis, by harnessing the power of machine learning, we employed the neural network potential model (NN) to predict the potential energy of HEA. The random arrangement of atoms makes training very challenging. In this thesis, by computing the material properties of CoCrFeNi HEA with the trained NN potential, we demonstrated that the NN potential energy predictor can successfully reproduce results from first principle calculations with errors around 1-5%. Nonetheless, the accuracy of NN energy predictors critically rely on the selection training sets. The NN potential model offers a hundreds of thousand times computational speedup than first principle calculations while retaining the computational accuracies, thereby allowing atomistic Monte Carlo method and molecular dynamics simulations to explore the configurations and chemical short-range orders of high entropy alloys.