藉由選擇性蝕刻金銅矽漸層膜與機器學習辨識爆裂物之拉曼光譜達成系統化之材料研究

Achieve systematic materials research by dealloying Au-Cu-Si gradient film library and identifying Raman spectra of explosives by machine-learning-based methods

方訓文(Hsun-Wen Fang)

指導教授 ： 薛承輝

國立臺灣大學/工學院/材料科學與工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201904336

摘要

本論文包含兩個主題。第一個主題旨在研究包含大範圍合金成分與結晶度分佈之金銅矽合金膜之選擇性蝕刻。第二個主題旨在藉由在類神經網路中加入生成式對抗網路以增進其分辨2,4,6-三硝基甲苯與2,4-二硝基甲苯之拉曼光譜的能力。其中，前者為最常見之爆裂物、後者則為前者的前驅物。本摘要之第二段及第三段將細述第一個主題，而第二個主題之細述則在第四段及第五段。
因其便宜、高產能、可大面積製作之特性，選擇性蝕刻一直是很盛行的奈米孔隙金屬材製程。當受到選擇性蝕刻時，合金中反應性最高的元素會被溶出，剩下的部分則會重組而形成新的合金形貌。此形貌中常有微米級至奈米級的孔洞。在選擇性蝕刻的影響因素中，母合金之成分及結晶度不但是合金的內在性質、且對選擇性蝕刻有巨大的影響力。故，針對這兩者研究之文獻很多。然而，這兩者之影響總是被分開討論。換句話說，兩者的共同效應常被忽視。
本研究利用裝備有金、銅、矽三個元素靶之磁控濺鍍機共鍍出具有成分漸層之金銅矽合金膜。而因結晶度與成分高度相關，此成分漸層直接導致了結晶度漸層的存在。藉由選擇性蝕刻此合金膜，可以製備出擁有不同成分及結晶度之金銅矽合金之選擇性蝕刻形貌。在此成果的幫助下，我們不但應證了一些先前的理論、也提出了新的發現。
拉曼光譜因其辨識化學結構之高度特定性而被稱作化合物之光學指紋。但因其產生機率很低，拉曼光譜很容易受雜訊影響。而由於在進行爆裂物的拉曼偵測時所需要的安全距離，此問題在爆裂物的拉曼偵測中會被進一步放大。
本研究中，我們嘗試以類神經網路以及其衍生之模型來分辨2,4,6-三硝基甲苯與2,4-二硝基甲苯之拉曼光譜。而由於最單純的類神經網路模型之表現受限於過度擬合，我們嘗試在訓練中加入一些不屬於前述兩者的資料。在這些資料中，只有在生成式對抗網路中的生成者所仿造之贗品能藉由其不斷演化的特性來增益模型表現。此效果被歸因於生成者與辨識者的對抗。而此對抗所引入之擾動可以幫助模型克服過度擬合並藉此增益模型表現。
這兩個主題看似無交集，其實共享一個目標：追求材料研究之效率。在第一個主題中，此目標體現於共鍍法的使用。藉由共鍍法，具備大範圍成分及結晶度的金銅矽合金群之製備可濃縮於一道製程。在第二個主題中，此目標則體現於生成式對抗網路之使用。使用生成式對抗網路解決在科學應用中常見之過度擬合可以避免使用啟發式且麻煩的超參數嘗試，進而提升研究效率。

關鍵字

共鍍 ； 選擇性蝕刻 ； 成分及結晶度 ； 奈米級孔隙 ； 拉曼光譜 ； 生成式對抗網路 ； 過度擬合

並列摘要

Two main topics are discussed in this thesis. The first topic aims at the dealloying behavior of the Au-Cu-Si gradient film library containing alloys with a wide variety of composition and crystallinity. The second topic eyes on accurately identifying the Raman spectra between TNT, the most prevalent explosives, and 2,4-DNT, a precursor of the former by introducing generative adversarial network (GAN) to a neural network (NN) classifier. Further details of the first topic is described in the 2nd and 3rd paragraphs while those of the second topic is described in the 4th and 5th paragraph of this abstract.
Dealloying has been a prevalent technique to produce nanoporous metals due to its cost-effectiveness, high productivity, and scalability. While dealloyed, the most reactive element in an alloy is selectively dissolved and the remains form new morphologies with pores sized from nanometers to micrometers. Among the factors influencing dealloying mechanisms, the composition and crystallinity of mother alloy have been studied extensively due to their intrinsic nature and significant influence. However, these two factors have always been studied separately. Namely, the combined effects have been overlooked.
In this study, by implementing co-sputtering with a magnetron sputter equipped with three targets (Au, Cu, and Si), a Au-Cu-Si film with a composition gradient was fabricated. This composition gradient would inevitably induce a crystallinity gradient as crystallinity depends on the composition. By dealloying the Au-Cu-Si gradient film, the dealloyed morphologies of various compositions and crystallinities were achieved. Enabled by these results, not only did some previously established theories were validated but also new ones were proposed.
Raman spectrum is known as the optical fingerprint of chemical due to its specificity in recording the chemical structures. However, the small scattering cross section of Raman scattering makes Raman detection vulnerable to noises. This problem even deteriorates in the detection of explosives where a standoff distance is required for safety concern.
In this study, we tried to classify the Raman spectra of TNT and 2,4-DNT, by some neural network (NN) models. As the most naïve NN model suffered from overfitting, the data didn’t belong to the Raman spectra of TNT and 2,4 DNT were also involved in the training. Among them, only the ever-evolving counterfeits generated by generators in GAN were beneficial to the performance of the model. The effect was later attributed to the competence between the discriminator and generator. Such competence, introduced additional perturbations to the model, helped the model to circumvent the overfitting, therefore, enhance the performance.
Seemingly independent, these two topics shared a common spirit of pursuing efficiency in materials’ research. While this spirit was embodied in the first topic by the co-sputtering where Au-Cu-Si alloys with a wide variety of composition and crystallinity was fabricated within one step, such spirit was realized in the second topic by the involvement of GAN who solved the overfitting problem, which is common in scientific applications, without the heuristic and troublesome hyperparameter searching.

並列關鍵字

Co-sputtering ； Dealloying ； Crystallinity and Composition ； Nanoporous ； Raman spectra ； Generative adversarial network ； Overfitting

參考文獻

1 Jin, H. J. & Weissmüller, J. Bulk nanoporous metal for actuation. Advanced engineering materials 12, 714-723 (2010).

2 Jin, H.-J. et al. Nanoporous Au− Pt alloys as large strain electrochemical actuators. Nano letters 10, 187-194 (2009).

3 Yun, Q. et al. Chemical dealloying derived 3D porous current collector for Li metal anodes. Advanced Materials 28, 6932-6939 (2016).

4 Meduri, P., Pendyala, C., Kumar, V., Sumanasekera, G. U. & Sunkara, M. K. Hybrid tin oxide nanowires as stable and high capacity anodes for Li-ion batteries. Nano letters 9, 612-616 (2009).

5 Yu, J. et al. Nanoporous metals by dealloying multicomponent metallic glasses. Chemistry of Materials 20, 4548-4550 (2008).