大會演講 Speakers

近來最熱門的討論議題之一，乃是人工智慧蓬(AI)勃發展之下，程式的撰寫(coding)需求，已經快速轉為AI在各領域的應用，其中跨領域的應用更成為學術、工業界中的重中之重。本講演將以我自身的經驗，從過去40餘年運用「台灣沒用的不做，沒有學術價值的不准做。」運用實際研究經驗來說明與檢驗這個跨領域的風潮與要求。

講演將從美國Cornell University畢業後，進入美國IBM Almaden研究中心的各項研究工作中，是用全像光學方法解決一項年營收高達20億美金產品的結構應力問題。當時如果無法及時處理，該產品每天的損失高達500萬美元。運用傳統材料科學的方法，需要耗費多年的時間才能解決這樣的問題，但產品遇到危機，僅有幾周解決問題，因此運用光學量測結合數學模型，在極短時間內提出創新的解法，順利的解決危機。所得研究成果因此被張貼在 IBM Research 部門的入口牆上，與 IBM Fellow 並列。「這不是傳統材料人的做法，但如果你用跨領域的方法去檢驗看、去解決，就可以快幾百倍。」

這樣的案例體現了美國國會近年提出的一項創新觀點：創新（innovation）並不是從基礎研究線性推進到應用，而是各階段高度互動、彼此回饋的結果。這也顯示，研究若能跨域學習、了解系統需求，將有機會在產業中扮演極關鍵的角色。

跨界複製：從基本理論到高階應用的潛力

IBM 對研發貢獻的獎勵制度。他回憶，自己每三至四個月就有機會獲得一次5％的薪資調整，遠超過傳統公司一年一次的固定加薪制度。他說：「你如果三年加薪 12%，那其實只是每年 4%，但 IBM 的做法是三、四個月就給你5％，長期下來加薪速度非常快。」加薪通知的方式也極具儀式感：只是一張寫在筆記本邊角的小紙條，寫著「Congratulations」。他笑說：「第一次收到還不知道是什麼，後來才知道是加薪通知。」

這樣的制度背後，是 IBM 強調「用成果說話」的文化。研究人員需要在短時間內提出可被實驗與市場驗證的解法，同時還要回溯問題的基礎理論，才能說服上層管理與決策者。他強調，這種高壓但有效率的工作模式，要求研究人員具備應用能力、理論深度與團隊合作精神的完美結合。

跨領域創新的思維實踐：從顯示器、風扇到羊毛

其次分享多個看似日常，實則蘊含高深技術的案例，進一步說明跨領域思維如何為產業帶來創新。

其中一例是 LCD 背光板的「光學墨水」問題。為了避免光線均勻性不佳，業者原本使用日本進口的特殊油墨進行印刷，卻因解析度提升導致良率大幅下降。李教授只花五分鐘就提出建議，改用特定尺寸的白色粒子（不超過 50 微米）混入透明膠體中，並控制比例不超過 50%，即可有效解決視覺均勻性與印刷良率問題。這個方案的材料成本僅約十元，但解決的是市面上售價近兩千元的顯示器背光問題，顯示數學與物理直覺如何在實務中產生驚人價值。

另一個令人印象深刻的案例是電風扇。他指出，臺灣市售的高階 DC 電風扇價格上看兩萬元，原因在於其特殊風葉設計能夠以渦流方式產生柔順、環繞的氣流，讓吹風不再是刺耳、直線的感覺，而是如「空氣在流動」。這樣的設計，不只是力學問題，也是數學與感知設計的結合。

最後一個來自紐西蘭的故事，則結合了生物科技與勞工福利。他描述紐西蘭科研機構如何為剪羊毛的工人解決脊椎側彎的風險，透過設計一種蛋白質注射法，使羊毛在特定時間點自然脆化，剪羊毛變得更輕鬆也更省力。這樣的設計兼顧數學建模、生物時間控制與社會問題回應，正是「real problem, real solution」的典範。

最終運用自己多次面對高風險、高回報研發任務的經驗進行結論：我們進行跨領域研究工作，其成效在AI的時代，乃是一個不可忽視的方法。

Computational hemodynamics has evolved from a research tool for understanding blood flow phenomena into a promising technology for clinical decision-making. Advances in medical imaging, computational modeling, and high-performance computing have enabled patient-specific modeling of cerebral circulation, providing new insights into cerebrovascular circulation. The increasing availability of large-scale clinical datasets further facilitates the integration of simulations with AI and data-driven approaches.

This lecture highlights recent developments in patient-specific computational hemodynamics by integrating medical imaging, multiscale cerebral circulation modeling, and large-scale data analysis. It will also discuss the emerging concept of a cerebrovascular digital twin, which combines image-based hemodynamic simulations with clinical data to support personalized prediction, risk assessment, and treatment planning.

Internal musculoskeletal loads, e.g., the muscle, ligament, and joint-contact forces that are central to orthopaedic function, injury, and disease, cannot be measured noninvasively in vivo owing to ethical or technological limitations. They must instead be inferred by combining experimental measurements with computational modelling: experiments supply the measurable inputs, namely subject-specific anatomy, joint motion, and external loading, while computation recovers the internal forces that no sensor can measure. However, each contributing discipline has tended to capture only part of this picture. Clinical orthopaedic assessment largely focuses on morphology, whereas conventional motion analysis measures only rigid-body kinematics. Computational or mathematical mechanical models can estimate internal loads but have been constrained by a lack of in vivo, subject-specific inputs, relying instead on generic or idealised geometry and motion. This lecture traces a combined experimental and computational pipeline that turns still medical images into living, load-bearing models, and argues that artificial intelligence is the enabling layer that makes these models subject-specific, accurate, and clinically deployable at scale.

The central aim has been to integrate three facets of the musculoskeletal system that are conventionally examined in isolation, namely its morphology, its motion, and its function, or mechanics, within a single subject-specific framework, and the resulting pipeline is presented as a progression from pixels to forces. Deep-learning segmentation and statistical shape modelling reconstruct subject-specific bone and joint geometry from CT, MRI, and even single planar radiographs; AI-accelerated model-to-image registration of fluoroscopic sequences recovers three-dimensional joint motion to sub-millimetre and sub-degree accuracy; and these geometries and motions drive subject-specific musculoskeletal and finite-element models that estimate the muscle, ligament, and joint-contact forces underlying normal and pathological movement. To the authors’ best knowledge, the integration of these components into a single image-to-force workflow represents a distinct departure from the generic, scaled-model approaches that dominate the field.

Applications across joint degeneration, surgical reconstruction, and sports injury illustrate how computational mechanics, so framed, converts routine clinical images into actionable mechanical insight. It is hoped that this image-driven, AI-enabled approach will help bridge computational mechanics and translational orthopaedic care.

Recent developments in computer-aided analysis and design methods have greatly simplified the creation of DNA origami nanostructures. In this talk, I will present our efforts to accelerate the design process of DNA origami by integrating finite element–based structural modeling with molecular dynamics–based characterization of their geometric and mechanical properties. Our computational framework enables rapid prediction of equilibrium configurations, thermal fluctuations, and dynamic reconfigurations of DNA origami nanostructures at molecular resolution.

In addition, I will introduce a generative design framework for DNA origami with completely free-form geometries based on diffusion models. The model is trained on simulated data generated using our coarse-grained modeling framework. By learning to translate arbitrary shapes into valid DNA structures, it enables the creation of novel forms unconstrained by traditional lattice-based design paradigms. This flexibility has facilitated the fabrication of a wide variety of structures, including free-form designs derived from sketches, pictograms, logos, and raster images. The framework also supports structural modularization and reconfiguration, opening the door to dynamic and adaptive nanostructures. Together, these advances represent a significant step toward expanding the expressive and functional design space of DNA origami beyond conventional constraints.

Inverse problems are central to engineering, from identifying material properties from sparse measurements to designing structures with prescribed mechanical responses. In this talk, I will present our recent work on physics-informed neural networks (PINNs) for material characterization, constitutive model discovery, and data-free inverse design of bistable kirigami and origami metamaterials. These studies demonstrate how governing physics can enable accurate inference and design even when experimental data are scarce or noisy. I will also discuss the limitations of PINNs and the importance of choosing how physics is incorporated into the inverse problem. Using pavement modulus backcalculation as an example, I will compare PINNs with differentiable finite element methods, which enforce governing physics directly through a differentiable solver. Together, these studies highlight a broader view of physics-informed AI that combines learning, automatic differentiation, and numerical solvers for robust and data-efficient engineering inverse problems.

Boiling and condensation are strongly influenced by the interaction between fluid dynamics, interfacial transport, and geometric confinement. Although significant efforts have been devoted to improving phase-change heat transfer, the fundamental role of geometry in governing flow structures and transport mechanisms remains incompletely understood. This presentation discusses recent computational studies on geometry-controlled phase-change heat transfer, with particular emphasis on boiling and condensation in confined channels and thermal management devices. Numerical simulations are employed to investigate how channel shape, curvature, confinement, and surface configuration influence bubble dynamics, liquid-vapor distribution, pressure characteristics, and heat transfer performance. Representative examples from microchannel boiling, condensation flows, and electronics cooling applications are presented to illustrate the underlying transport physics. The talk further discusses how computational mechanics can be utilized to identify geometry-performance relationships and provide design guidelines for next-generation thermal management systems.