以機器學習方法訓練四元過渡金屬硫屬化合物二維材料之分子勢場

Abstract

二維過渡金屬二硫屬化合物 (2D Transition Metal Dichalcogenides, TMDs) 是一種結構類似於石墨烯 (Graphene) 的化合物，其結構特徵為在同平面上原子排序呈現六邊形蜂巢狀，且側面原子間排序呈現上下交錯排列。金屬(如二硫化鈮、二硫化鉭)或半導體(如二硫化鉬、二硒化鎢)等二維材料相繼問世後廣泛應用在光學、光電探測器上，因其熱穩定性高，可在室溫下穩定存在於大氣中，使得此類二維材料迅速成為熱門材料及研究的熱門課題。 欲進行多元金屬硫屬化合物二維材料之理論計算來探討不同濃度結構的混合能量以及基態能量，由於其化學環境複雜，難以找到準確的古典勢能場以進行分子動力學模擬；第一原理計算雖能精確計算原子間作用力，但需要耗費大量的計算資源及時間，且計算尺度被限縮在幾百個原子內。在本研究中，考慮到過渡金屬硫屬化合物不同原子的隨機排列，系統規模超過第一原理負荷。因此，本研究使用第一原理計算之小尺度系統當作訓練範本，結合神經網路與機器學習方式擬合出能夠準確描述大尺度系統並進行分子動力學的勢能模型。 訓練完成的神經網路勢能先以測試集驗證其系統能量與各別原子受力之準確度，顯示出神經網路勢能除了有與第一原理相當之準確度，且具有更高的計算效率，亦可應用於第一原理難以處理之大系統。神經網路勢能模型能夠迅速又準確的估算複雜的化學環境，進而進行大尺度的模擬以探索其材料性質。本研究以約兩千個原子的結構進行蒙地卡羅模擬，在不同種多元素組成的結構下計算不同濃度結構的混合能量以及基態能量，其複雜的化學結構很難從實驗中獲得，機器學習勢能模型展示出可以用來探索微觀結構下的複雜化學環境。

2D Transition Metal Dichalcogenides (TMDs) is a compound material with a structure similar to graphene. The structures of TMDs feature the sequence of the coplanar atoms with a hexagonal honeycomb shape similar to graphene, and the sequence of staggered up and down side atoms. Recently, metallic dichalcogenides (NbS2, TaS2) or semiconducting dichalcogenides (MoS2, WSe2), have been widely used in the optical and photodetectors applications. Because of their high thermal stability, most TMD materials are stable under ambient conditions, making them become a popular topic in advanced materials. Recently mixed TMDs such as (MoxW1-x)(SySe1-y)2 are drawing an increasing attentions owing to highly tunable properties by changing composition x and y. The mixing energies of this mixed TMD material as the function of chemical compositions provide critical microstructural information of the material. Nevertheless, such information can only be extracted from theoretical calculations. The first principle calculations can evaluate the system energetics of such chemically complex materials with arbitrary compositions; however, the first principle calculations are computationally intensive, making exploration of the configurational space of mixed TMDs infeasible. In this thesis, we harnessed the power of machine learning by training an artificial neural network (ANN) potential model. The ANN model was trained by generating a training set comprised of atomistic images of (MoxW1-x)(SySe1-y)2 mixed TMDs of different compositions, labeled by energies computed from the first principle calculations. Once successfully trained, the ANN model can be utilized for large-scale atomistic simulation of mixed TMD materials of arbitrary size beyond the reach of the first principle calculations. We demonstrated that the trained ANN potential model can predict the energies of TMD material with high fidelity to the first principle calculations; furthermore, the ANN model also offers orders of magnitude computational speed up relative to the first principle calculations, thereby allowing us to perform exhaustive sampling over the configurational space of the mixed TMD materials. We performed Monte Carlo simulations of the (MoxW1-x)(SySe1-y)2 mixed TMD material of different compositions, with a system size of around two thousand atoms —- a size literally impossible for the first principle calculations. We computed respective mixing energies, demonstrating that the machine-learning-enabled energy model can be utilized to explore the microstructures of chemically complex materials that are difficult to be retrieved from experiments.