機器學習輔助低楊氏係數鈦合金開發

Abstract

為了解決傳統生醫合金所擁有的應力不匹配及低生物相容性的問題，許多新的醫用鈦合金成分不斷地被開發出來，以取代舊有合金成為新的骨替代材。然而，目前主流的合金設計方式導致了新的合金成分被限制在某些特定的合金成分範圍。這樣的合金範圍不僅使得合金之原料價格增加，也產生在熔煉及合金製作上許多困難。因此，本論文即聚焦於改善前述之合金設計缺點，致力於利用機器學習的方式，發展新的合金預測系統，進行合金開發。此論文將會橫跨機器學習與材料製造之不同領域，從各種不同的面向開發新的候選合金。 在第四章中，此論文開發了一個可以協助低楊氏係數β相鈦合金成分開發之合金選擇工具稱作βLow。此工具包含兩個獨立之類神經網絡，分別預測不同合金成分之β相穩定性及彈性模數。此工具不僅僅是目前少數能夠提供彈性模數預測之方法，其在β相穩定性的表現，更是超越過去主流之β相預測方法。在第五章中，此論文透過開發更新版本的βLow 2.0來更深入的討論機器學習對於合金設計的看法。 βLow 2.0 比起前一代的機器採用新演算法及更大的資料庫。從結果來看，此更新提升了預測表現，讓其在相穩定性及楊氏係數的預測上皆成為目前表現最好的模型。透過分析輸入輸出參數之間的關係及預測不確定性之量化，此研究從βLow 2.0的預測出發，提出了四個對於未來合金設計的觀點，其中包含了相穩定性、β穩定元素添加、中性元素添加及機器學習輔助合金設計方向建議。 在第六章中，透過機器的推薦，本論文發現了一個前所未見的全新鈦合金成分, Ti-12Nb-12Zr-12Sn (12Nb)，其不僅能夠有低的楊氏係數，也擁有低合金價格。在此章節中，透過顯微結構觀察、機械性質量測及生物相容性測試發現此合金不僅擁有低的楊氏係數，也有著非常好的機械性質、擬彈性性質及生物相容性，是未來骨質入材料的人選之一。第七章則是聚焦於如何透過墨水擠出列印製作12Nb及Ti-6Nb-6Mo-12Zr-12Sn (6Nb6Mo)。此列印方法的特點在於其能夠透過元素粉末的混合形成合金，提供了合金設計上的自由度。此外，此研究的特色在於12Nb及6Nb6Mo擁有著相似的合金成分，卻擁有著完全不同的相穩定性及機械性質，也對於列印後的顯微結構與機械性質有所影響。此章節先利用背向散射影像、能量色散X射線譜及電子背向散射繞射研究此兩種合金在燒結及均質化過程的顯微結構變化。接著，透過micro-lattices的壓縮測試，我們得到6Nb6Mo有著相較於12Nb較高楊氏係數及降伏應力之特點。在第八章中，利用前述墨水擠出列印方法的優勢，將12Nb及6Nb6Mo合金以各半相疊及層狀相疊的方式堆疊成混合micro-lattices。如同擴散係數之預測， Nb及Mo的交互擴散僅發生在連接處的第一層，同時交互擴散的結果抑制了12Nb中的α相生成。經過熱處理之後，如同前一章的結果，兩合金最後都呈現純β相之顯微結構。在機械性質的方面，壓縮測試的結果顯示堆疊方式對於機械性質有顯著的影響。層狀相疊有著更好的機械性質，包括6Nb6Mo的高強度及12Nb的低楊氏係數。 此博士論文旨在為機器學習協助合金開發之議題提供一個嶄新的例子，包含模型訓練、合金尋找、顯微結構控制及產品製造。期待本研究所提供的知識及討論的內容，將會加速骨替代材之開發進程，也同時能被應用在其他產品之開發。

Dealing with the problems of stress shielding effect and low biocompatibility in conventional alloys, which mainly used for biomedical applications, many new biomedical titanium alloys with low modulus were developed. However, compositions of newly discovered alloys were limited in certain intervals by following the mainstream material design methods. The limitation not only increases price of raw material, but also increases difficulties in melting and producing products. Therefore, this work focuses on applying machine learning in alloy design to improve the weakness of above-mentioned material design methods, endeavoring breaking current composition limitations for titanium alloys. This work covers different aspects in material design and manufacturing methods, including machine learning, thermomechanical processing, and additive manufacturing. In Chapter 4, a new alloy selection tool, named βLow, is developed for low modulus β phase titanium discovery. It consists of two individual artificial neural networks for Ms temperature and modulus prediction. βLow is not only a rare prediction tool that provide modulus prediction, but also a tool that have solid prediction in phase stability, which is better than mainstream methods. In Chapter 5, a new version of tool, βLow 2.0 is developed with the desire of understanding how models work. βLow 2.0 is trained by improved algorithm and expanded datasets, and is proven to be state-of-art machine in both stability and modulus prediction. By conducting sensitivity analysis and uncertainty quantification, four insights from βLow 2.0 are revealed for future alloy design, including effect of phase stability, effect of β stabilizer addition, effect of neutral element addition, and material design direction. In chapter 6, a basic analysis is conducted on Ti-12Nb-12Zr-12Sn (12Nb) alloy. This alloy is discovered from βLow and features low modulus and low cost. The result further suggests that this alloy has good mechanical properties, pseudoelasticiy, and biocompatibility. Therefore, it holds promise for future application in orthopedic implants. In chapter 7, 12Nb and Ti-6Nb-6Mo-12Zr-12Sn (6Nb6Mo) is manufactured by ink extrusion printing based on elemental and hydride powders. The two alloys have similar compositions but extremely different mechanical properties and phase stabilities. The microstructure evolution during sintering and homogenization processes is studied via backscattered electron microscopy, energy dispersion X-ray spectra, and electron backscattering diffraction on printed filaments. Micro-lattices with single β phase are tested in compression, with 6Nb6Mo showing higher modulus, ductility, and yield stress than 12 Nb. In chapter 8, the above-mentioned two alloys are further stack-printed as hybrid micro-lattice structures, with 12Nb and 6Nb6Mo layers interspersed at the layer level or as two separate blocks. As predicted by diffusion coefficient of Nb and Mo, inter-diffusion between two elements only occurs at the contacted layer. The diffusion of Mo from 6Nb6Mo limits the formation of α phase after sintering. After phase homogenization, both sides of alloys obtain single β phase in microstructure. The two stacking strategies lead to different compressive properties of scaffolds, with the multi-layered hybrid exhibiting a desirable combination of good strength from 6Nb6Mo and low stiffness from 12Nb. This dissertation aims at a new view in machine learning assisted material design, including details for each step – model training, alloy discovery, microstructure control, and manufacturing. The topics and information in this dissertation are expected to accelerate the development progress of orthopedic implant and even apply to the development of other products.