Articular cartilage is a crucial tissue between the ends of bones that provides protection to the joint through its deformation-resistant and load-bearing ability. If cartilage is subjected to abnormal mechanical stimulation, such as impacts or repetitive loading, excessive degradation of the cartilage extracellular matrix may occur. This results in the degeneration of the cartilage and the development of related diseases, such as osteoarthritis. A disintegrin-metalloproteinase with thrombospondin motifs-5 (ADAMTS-5) is an important peptidase which catalyzes aggrecan, the primary proteoglycan in the cartilage extracellular matrix. Intriguingly, it is found that there is a non-cleavage site Glu419-420Ala in the interglobular domain of aggrecan, which is composed of the same amino acids as the critical cleavage site Glu373–374Ala of ADAMTS-5. In this research, we introduce an innovative bottom-up in silico approach to reveal the substrate recognition mechanism of ADAMTS-5 from an atomic-scale perspective. It is found that besides the actual cleavage site, the P7 and P10’ residues also play a critical role in the recognition and catalytic mechanism due to their better binding affinity. The three-point (P7, P10’, and the cleavage site) binding mode enables the actual cleavage site of aggrecan to bind to ADAMTS-5 more stably and firmly. Understanding the molecular mechanism of extracellular matrix degradation paves the way for future medical and pharmaceutical research of prophylaxis and treatment for related diseases.

keywords: extracellular matrix, aggrecan, ADAMTS, multiscale modeling, molecular dynamics simulations

鎳鈦形狀記憶合金(Shape Memory Alloys)擁有傑出的機械、熱耦合性質，在受到外在環境刺激時，能表現出不尋常的特性。例如：形狀記憶效應與超彈性現象，這些超常的性質讓鎳鈦形狀記憶合金被廣泛應用於致動器與阻尼器等機械元件中。為了瞭解造成這些現象的成因，本研究從微觀尺度下麻田散體相(Martensite)兄弟晶相變化機制作為基礎，利用超級電腦的計算力，將模擬的空間尺度推進到介觀尺度的規模，探討微結構的變化與巨觀表徵的關聯性。
本研究使用分子動力學方法，模擬鎳鈦形狀記憶合金的壓印過程。壓印完成後再進行定溫與升溫模擬，藉此觀察形狀記憶效應所對應之微結構演化過程，進而了解單向或雙向形狀記憶效應的產生機制。微結構辨識方面，除了採用傳統的共鄰原子分析(Common Neighbor Analysis, CNA)之外，本團隊亦開發了基於圖神經網路(Graph Neural Networks)的麻田散體兄弟晶辨認法(Martensite Variant Identification)，可以得到相變化發生的時機以及更完整的微結構辨識，進一步分析其對機械性質的影響。在放大模擬尺度方面，本團隊利用中小型壓印模型的經驗資料與經驗，搭配國網中心的台灣杉三號超級電腦，放大了模擬空間，消除了邊界效應與尺寸效應可能帶來的影響。受限於計算資源的取得，其他相關研究文獻著墨甚少，本研究完成的尺度效應測試更顯得得來不易。
為了完全發揮硬體的性能並且盡可能加大模擬的系統，我們將原先能在台灣杉一號運算的奈米級壓印模型(40 nm * 40 nm * 30 nm)，針對台灣杉三號CPU架構重新規畫空間分散策略，在台灣杉三號alpha test階段，進行了強弱耦合的效能評估(strong / weak scalability)。並於測試期間，以720 nodes進行超大型壓印模擬，完成尺寸達250 nm * 250 nm * 210 nm，共2.9億顆原子，壓印深度達9nm的模擬。
面對如此龐大的計算結果(24.5TB)，過去使用單一個人電腦進行分析已不可行，我們使用了分散式計算方法，將分析工作依空間關係切割為數千個獨立小工作，再以台灣杉三號的計算資源批次處理。而視覺化的呈現，我們也利用開源的科學視算軟體paraview進行平行繪圖(parallel rendering)，將微結構的分類結果進行視覺化呈現。本研究中放大了模擬空間，消除了模擬邊界與尺寸效應對於壓印結果的影響，對於微觀行為的模擬視至關重要的。而為了如此大規模的模擬所發展之平行處理、分析及視覺化技術，也將對國內計算模擬研究群帶來相當助益。

Biomaterials exhibit extraordinary properties, partly due to intricate structures at the microscale. Leaning from these microstructures is essential to design high-performance structural materials. Therefore, in this research, we combine the idea of AutoEncoder, Transformer, and Generative Adversarial Network (GAN) to create AE-Transformer-GAN, a new framework that can learn the information from cross-sectional slices of biological microstructures and generate the biological-like 2D images that can later be assembled back into the 3D bioinspired microstructures. We first use our proposed framework to generate Gyroid, a mathematically well-defined microstructure with a strong periodicity. We then demonstrate the robustness of the framework by generating Elk Antler, a biological microstructure with marvelous axial compressive strength. To verify the effectiveness of our proposed framework, we examine the similarity between expected slices and generated slices. Moreover, we compare the continuity between the expected sequence and generated sequence to evaluate the capability of our proposed framework to process sequential data. All results suggest that the model has remarkable performance on learning patterns and generating microstructures. In conclusion, our proposed framework, AE-Transformer-GAN, can serve as an feasible way to learn and generate de novo biological 3D microstructures in a reasonable effort.

The study of natural and bio-inspired materials with their extraodirnary properties are rapidly emerging. Most of these properties are a direct consequence of microstructures. Thanks to the advancement of high resolution computed tomography and 3D printing, the inherient complexity of microstructures from biological materials are now accessible and producible. However, since these biological microstructures are not originally designed for engineering applications, such microstructures may not faultlessly contend with our requirements. Therefore, the objective of this research is to propose a method that can mix different biological microstructures and create unprecedented bio-inspired microstructures with properties beyond the ordinary mixing rule. In this research, we adapt the idea of Auto-Encoder and the framework of Style-Based Generative Adversarial Networks (Style-GANs) and propose AE-Style-GAN, a deep learning model that can combine the advantages of the referred samples and generate new microstructures without their innate defects. To test the capability of AE-Style-GAN, we apply our proposed model to mix Gyroid and Antler, two biological materials with complementary mechanical properties. We then perform projection and similarity analysis on the mixed microstructures to examine the geometry correlation between the referred and mixed microstructures. Furthermore, we construct the actual samples of our mixed microstructures with a 3D printer and conduct a compression test on the printed samples to verify its mechanical properties. The results suggest that our model is capable of mixing different biological materials and synthesizing microstructures with properties beyond the mixing rule. In conclusion, we expect this method can be exploited to mix various biological materials and synthesize more potential and unparalleled microstructures for engineering applications.

Their folding structure can determine the function of proteins. Therefore, the ratio of alpha-helix and beta-sheet in secondary structure is a key to design the functional protein we desired. To this end, we propose a deep generative model that can generate proteins with a specified ratio of the secondary structures. Given a specific range of 2nd structures, the model can create protein sequences that meet our requirements. Our model was composed of a deep learning model to analyze the ratio of 2nd structures and a genetic algorithm to perform inverse design. The model was trained by more than a hundred thousand sequences of proteins from Protein Data Bank (PDB) . The training results exhibited outstanding performance on determining the ratio of specific 2nd structure with an accuracy over 99%.
Moreover, the testing results showed promising 2nd structures contents, which we desired outside our training dataset. We further compared the sequences with the PDB dataset to verify their uniqueness. The comparison indicates that the generated sequences have only 50%~70% similarity with existing sequences. These results showed that we could control the ratio of alpha-helix and beta-sheet to generate proteins by design. This model can be applied to drug discovery, the structure design of tissues, and used for antibody and vaccine development.
Keywords: function of proteins, secondary structure, biomaterials, artificial intelligence

Fracture behaviors of brittle materials have been a crucial problem when it comes to safety and reliability. With deep learning (DL) recently drawing attention in material design and nanoengineering, several studies have applied deep learning techniques to predict the crack growth of brittle materials. This study aims to use a convolutional LSTM model to predict the fracture path of graphene under two systems of defective configuration. One involves graphene with different crystalline, and the other involves graphene with different porous defects. First, we perform tensile test simulation on graphene using molecular dynamics (MD) simulation. The fracture results are then transformed into image-based data in gray scale for the deep learning model. Finally, we construct two ConvLSTM-based model to learn the spatial-temporal information about the crack propagation for those two different graphene systems, respectively. After using 100 crack images in the cases of system with different surface orientation, results show that our ConvLSTM-based models can predict the fracture path of graphene with 99 percent accuracy on a system of different crystalline. On the other hand, by using 295 crack images in the other case, results show that our model has the capability of predict fracture growth with 98 percent accuracy on porous graphene. These models demonstrate the power of exploiting deep learning on nanoengineering and confer desired properties of materials, which has great potential for next-generation materials by design.

The state-space representation is an alternative mathematical tool which is common in the system theory to describe system behavior instead of the well-known input-output representation. In this study, we use state-space representation to explore the mechanical behavior of the tricuspid valve (TV) chordae tendineae. To model the mechanical behavior of the TV chordae, we connect the generalized Kelvin model with an additional nonlinear spring to establish a nonlinear generalized Kelvin model (NGK model) and cast it into the state-space description. Furthermore, the exact solution to the constant force, the constant-rate force, and the cyclic force inputs are obtained. The exact formulations of the tangent compliance, the creep function, and the retardation time of the NGK model are derived. On the other hand, the response of the NGK model under displacement controls is solved with the aid of the state-space representation. The influence of initial condition of state variables on the model behavior is investigated. After the theoretical preparation, the comparison between the simulation of the NGK model and experimental data of chordae tendineae are made and shows accepted agreement.
keywords：chordae tendineae; state-space representation; viscoelastic model; the nonlinear generalized Kelvin model; displacement control.